Author’s Accepted Manuscript

Forecasting stock index futures returns with mixed-

frequency sentiment

Bin Gao, Chunpeng Yang

www.elsevier.com

PII: S1059-0560(17)30058-8
DOI: http://dx.doi.org/10.1016/j.iref.2017.01.020
Reference: REVECO1365
To appear in: International Review of Economics and Finance
Received date: 17 June 2015
Revised date: 13 September 2016
Accepted date: 16 January 2017
Cite this article as: Bin Gao and Chunpeng Yang, Forecasting stock index futures
returns with mixed-frequency sentiment, International Review of Economics and
Finance, http://dx.doi.org/10.1016/j.iref.2017.01.020
This is a PDF file of an unedited manuscript that has been accepted for
publication. As a service to our customers we are providing this early version of
the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting galley proof before it is published in its final citable form.
Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal pertain.
 Forecasting stock index futures returns with mixed-frequency sentiment

Bin GAO1,2， Chunpeng YANG3*

1
School of Business Administration, South China University of Technology, Guangzhou, China
2
College of Economics, Qingdao University, Qingdao, China
3
Financial and Securities Research Center, School of Economics and Commerce, South China
University of Technology, Guangzhou, China
*
corresponding author: Tel: +8613533645882. ycpgroup1@gmail.com

Abstract

Using the data in Chinese financial market, mixed-frequency stock index futures sentiment and

mixed-frequency stock index sentiment are constructed according to MIDAS Model. We test whether

mixed-frequency stock index futures sentiment and mixed-frequency stock index sentiment have predictive power

on stock index futures returns. The empirical results show that mixed-frequency stock index futures sentiment

factors have more predictive power than mixed-frequency stock index sentiment factors and Fama-French three

factors. In out-sample forecast, we show that sentiment trading strategy provides a more positive returns than

time series momentum trading strategy and passive long positions.

Keywords: Stock index futures sentiment; Stock index sentiment; Mixed-frequency; MIDAS Model;

1 Introduction

Recently, more and more analyses have discussed sentiment factors can affect returns in futures markets. This paper

test whether mixed-frequency sentiment factors have more power in predicting stock index futures returns by daily,

weekly, monthly frequencies. We find that mixed-frequency stock index futures sentiment and mixed-frequency stock

index sentiment have predictive power significantly in-sample test. In out-sample forecast, we find sentiment trading

strategies has a statistically significant return, a annualized return of 80% and annualized Sharpe ratio of 1.75.

Why might sentiment factors help to forecast movements in returns? Some researches of the sentiment-returns

1
relationship have found that sentiment is significant and systemic pricing factor, and sentiment could predict risk asset

returns. In sentiment-returns relationship researches of stock market, stock sentiment usually are sampled at quarterly

data; monthly data and weekly data in empirical researches, such as quarterly data [Cen et al. (2014)]; monthly data

[Brown and Cliff (2005), Baker and Wurgler (2006), Baker et al. (2012), Kim and Ha (2010), Kurov (2010), Kumar

and Lee (2006), Schmeling (2009), Verma and Soydemir (2009), Yu and Yuan (2011), Kim et al. (2014), Fong and

Toh (2014)] and weekly data [Joseph et al. (2011)]. In futures market sentiment-returns relationship researches,

futures sentiment usually are sampled at weekly data or daily data in empirical research, such as weekly data [Wang

(2001), Wang (2003)] and daily data [Simon and Wiggins III (2001), Safa and Maroney (2012)]. For example, Simon

and Wiggins III (2001) find futures sentiment have negative predictability on futures returns in long term (30-days

lags). Wang (2001, 2003) find that large speculator sentiment forecasts price continuations by weekly data. In contrast,

large hedger sentiment predicts price reversals. Traditional sentiment-return research ignored the fact that the

sentiment effect is different in the short-term and long-term investmen and cause inaccuracy of prediction.

The variables which are available at high frequency contain potentially valuable information and have better

predictive power than low frequency variables. See for example, some study [Andersen et al. (1999), Andersen

et al. (2003), Hallam and Olmo (2014a, b)] show that the use of higher frequency data can provide significant

improvements in the modeling and forecasting of the daily return volatility compared to models using only

daily data.

We consider that it is impotent to analyze the mix-frequencies of the sentiment variables. Irrational factors

have more effect on high frequency decision-making by the financial experiment of Mcclure et al. (2004). Yang

and Gao (2014) and Yang and Zhou (2015) find high frequency sentiment factors have greater effect on asset

returns in both spot stock index and stock index futures market. In the actual investment, investors consider not

only short-term investment strategies but also long-term investment strategies. The returns are affected not only

by low frequency sentiment factor but also by high frequency sentiment factor. Therefore, we use

mix-frequency sentiment as predictive repressor.

Our study focuses on three questions. First, do sentiment variables possess predictive power over stock index

futures returns? To address this issue, we explore predictability for a range of stock index futures price indexes from

April 19, 2010 to April 26, 2013 by using the data in Chinese stock index futures market. Some studies also have

2
examined the predictive power of sentiment factors on returns in the futures markets. Wang (2001) studies the

predictive power of a trader position-based sentiment index in six agricultural futures contracts. Simon and Wiggins

III (2001) examine the predictive power of market-based sentiment. Wang (2003) investigates whether and how

sentiment factor based on trader positions forecast future market movements in the S&P 500 index futures market.

However, the futures sentiment from the literature is single sentiment index based on trader-position sentiment. We

construct stock index futures sentiment according to first principal component method and find the composite

sentiment index is better than single sentiment index in predictability.

Second, how does stock index futures returns predictability vary with the data frequency? To address this question,

we consider the monthly, weekly, and daily data frequency. Most sentiment-returns researches ignored the data

frequency effect, their sentiment index usually are sampled at some single frequency in empirical researches. Some

literatures have cared to different data frequencies [Lee et al. (1991), Brown and Cliff (2005): weekly data and

monthly data; Wang et al. (2006): Greenwood and Shleifer (2014): daily data and weekly data; quarterly data,

monthly data and weekly data]. But the related literatures mainly focused on the consistency of different frequencies

and they did not pay attention to heterogeneous influences at the different frequencies.

Third, does stock index futures returns predictability increase based on mixed -frequency sentiment factors？

The MIDAS regressions have been applied in the financial market, such as the research of financial volatility

and mixed-frequency sentiment in stock market, in Yu and Yuan (2010) and Yang and Zhang (2014). Yang and

Zhang (2014) use quarterly/monthly/weekly trading data, and examine whether mixed-frequency investor

sentiment affects stock returns. They find that mixed-frequency investor sentiment is more important than the

low-frequency one. Clearly, it is of interest to see whether a similar finding carries over to stock index futures

markets. In this paper, we use monthly/weekly/daily trading data and test whether mixed-frequency stock index

futures sentiment and mixed-frequency stock index sentiment increase the predictive power on stock index futures

returns.

For testing the effect of mixed-frequency sentiment in MIDAS regressions, we find that (1) mixed-frequency

stock index futures sentiment and mixed-frequency stock index sentiment are statistically significant and they

have greater positively predictive power on stock index futures returns. (2) Mixed-frequency stock index

futures sentiment is more statistically significant and has greater positively predictive power on stock index

3
futures returns than mixed-frequency stock index sentiment. (3) Mixed-frequency stock index futures

sentiment and mixed-frequency stock index sentiment have greater positively predictive power in high

sentiment period. (4) Mixed-frequency sentiment factors have greater positively predictive power than

Fama-French factors.

In addition, we evaluate the out-of-sample forecasting performance of mix-frequency sentiment factors by

Root mean square error (RMSE). We also compare our results against a benchmark model (AR(1) model). By

comparing RMSE, we show the RMSE of mix-frequency sentiment predictions are smallest, so the

mix-frequency sentiment predictions perform better than low-frequency sentiment predictions and benchmark

model. Compared with the in-sample evidence, the out-of-sample results are robust that mixed-frequency

sentiment factors have greater predictive power on stock index futures market.

Our results are helpful for understanding the fact that sentiment factors have great effect on financial

decision-making. We find sentiment factors have strong predictive power on returns, specially, the predictive power

of monthly/daily mixed-frequency sentiment is best with a t-statistic of 20.33 and Q-test of 16.45. More important, we

find sentiment trading strategies works for returns significantly, an annualized return of 80% and annualized Sharpe

ratio of 1.8. In out-sample profitability analyze, we compare the profitability of sentiment trading strategies,

momentum trading strategies and passive long strategies. We show the performance of sentiment strategy provides a

relatively steady stream of positive returns that outperforms momentum trading strategy and passive long positions.

The remainder of the article is organized as follows. In section 2, we construct stock index futures sentiment

and stock index sentiment. Section 3 presents the methodology and describes the in-sample and out-of-sample

evidence empirical results from regression model. Section 4 presents evidence of sentiment trading strategies.

Section 5 adds the evidence of predict returns by the relation of sentiment and returns. A conclusion ends the article

in section 6.

2 Variables and Data

This section describes the data sources for the stock index futures returns and predictor variables (stock

4
index futures sentiment and stock index sentiment), and provides a brief characterization of the data.

2.1 Stock index futures returns

To further develop Chinese financial markets, the CSI 300 Index futures contract was launched on April 16, 2010

on the China Financial Futures Exchange. During the sample period (April 19, 2010 to September 30, 2014), the

volume/open interest ratio declined from the 26 times to the 8 times. However, the average volume/open interest

ratio is generally less than 1 in the overseas mature market; the open interest is even bigger than the volume 30% to

40% in the U.S futures market. The average daily open interest (47283 contracts) is smaller than average daily

volume (399226 contracts) from April 19, 2010 to September 30, 2014. The expiration day of the CSI 300 index

futures contact is the third Friday of the contract (delivery) month, and the contract (delivery) months include the

current month, the next month, and the final months of the next two quarters, which are called quarter-months.

Transaction codes are IFL0, IFL1, IFL2 and IFL3. The data is from the China Financial Futures Exchange

(http://www.cffex.com.cn).

The close-to-close daily returns of four contracts (IFL0, IFL1, IFL2, and IFL3) Ri ,t are calculated as follow:

Ri ,t  100  ln( Pi ,t / Pi ,t 1 ) i  1, 2,3, 4 (1)

2.2. Stock index sentiment

In this part, we set up daily/weekly/monthly CSI 300 stock index sentiment based on first principal component

method. It has been used in some empirical researches and theoretical models on stock index sentiment, such as Li

(2014) and Yang and Gao (2014).

Based on the related existing literatures, we consider four separate proxies: trading volume (TV), adjusting

turnover rate (ATR), buy–sell imbalance (BSI), and psychological line index (PSY).

Four stock index sentiment proxies are used in the empirical tests. Trading volume, or more generally liquidity,

can be viewed as a stock index sentiment index by Baker and Wurgler (2006). Baker and Stein (2004) note that

5
trading volume can stand for market liquidity; it could be as sentiment indicator. Yang and Gao (2014) use trading

volume as stock index sentiment proxies.

Baker and Stein (2004) suggest that the turnover rate can serve as a sentiment index. But the turnover rate can’t

show whether the investors are optimistic or pessimistic. Yang and Zhang (2014) suggest that use adjust turnover

rate instead of turnover rate as sentiment indicator. Yang and Gao (2014) also use adjusting turnover rate as stock

index sentiment proxies.

Specifically, the adjusted turnover rate ( ATR ) is as follows:

Rit VOLit
ATRit   .
Rit shares outstanding at time t

Where Rit is the return of stock index i at time t , VOLit is the trading volume of stock index i at time t . If
return is above zero, the adjusted turnover rate is positive to indicate that the stock market is bullish. If return is
below zero, the adjusted turnover rate is negative to indicate that the stock market is bearish.
The third proxy is based on buy–sell imbalance. Kumar and lee (2006) use the trading activities of retail investors

to measure changes in their sentiments. Yang and Gao (2014) consider buy–sell imbalance as stock index sentiment

proxies.

The primary data for the buy and sell imbalance ( BSI ) consist of the buyer-initiated volume ( BV ) and the
seller-initiated volume ( SV ). In this paper, we focus on the buy and sell imbalance as a parsimonious investor
trading behavior index in China: the buyer-initiated volume and the seller-initiated volume. To compute the buy and
sell imbalance index BSI , we define the day- t BSI spot for stock index as:

BVspot ,t  SVspot ,t
BSI spot ,t 
BVspot ,t  SVspot ,t

Here, BVspot ,t is the buyer-initiated volume of stock index on day t , SVspot ,t is the seller-initiated volume of

stock index on day t . A given day’s stock index BSI spot indicates whether most of the investors are buyers

( BSI spot  0 , that is, the buyer-initiated volume is greater than the seller-initiated volume) or sellers ( BSI spot  0 , that

is, the seller-initiated volume is greater than the buyer-initiated volume).

The fourth proxy is a futures sentiment index based on psychological line index. The Psychological Line is a

sentiment indicator by Kim and Ha (2010). Yang and Gao (2014) consider psychological line index as stock index

sentiment proxies.

The psychological line index ( PSY ) is as follows:

PSYt  T u T 100.

6
Where T u is the number of days when the closing price of stock index i at time t is higher than the closing
price of stock index i at time t  1 , and T is the trading period. And the market is overbought with an PSY of 75,
and the market is oversold with an PSY of 25.
The sample period ranges from April 16 2010 to April 26 2013. The data is from the RESSET database and WIND

database. The stock index sentiment proxies are measured daily, weekly and monthly.

A composite index is based on their first principal component. This procedure leads to the stock index sentiment

index , for example (e.g. monthly):

Sspot ,t  0.345TVt  0.454 ATRt  0.421BSIt  0.224PSYt

Each of the four components is standardized. The first principal component explains 75.41% of the (standardized)

sample variance, and only the first eigenvalue is far above 1.00, so we conclude that one factor captures the common

variation.

2.3. Stock index futures sentiment

In this part, we set up daily/weekly/monthly CSI 300 index futures sentiment as futures sentiment based on the first

principal component. It has been used in some empirical researches on stock index futures sentiment, such as Yang

and Gao (2014).

Based on the related existing literatures, we consider four separate proxies: trading volume (TV), open interest (OI),

buy–sell imbalance (BSI), and psychological line index (PSY).

Four futures sentiment proxies are used in the empirical tests. Baker and Stein (2004) note that trading volume can

stand for market liquidity; it could be as sentiment indicator. Scheinkman and Xiong (2003) note that volume reveals

underlying difference of opinion, which are in turn related to valuation levels when short selling is difficult. Yang and

Gao (2014) also use trading volume as stock index futures sentiment proxies.

The second proxy is a futures sentiment index based on open interest. By Wang (2001), open interest is the futures

sentiment index, rather than net positions or excess long (or short) positions, and is chosen to study returns

predictability in futures markets for the following reasons. Wang (2001) notes that open interest provides a

more-intuitive reading of trader actions, it is similar in nature to other sentiment indexes in the market place, and

7
widely accepted by futures participants. Yang and Gao (2014) also use open interest as stock index futures sentiment

proxies.

The third proxy is a futures sentiment index based on buy–sell imbalance. Wang et al. (2011) also use the buy-sell

imbalance as sentiment indicator from the Taiwan electronic stock index (TE) and the Taiwan financial stock index

(TF). Yang and Gao (2014) consider buy–sell imbalance as stock index futures sentiment proxies.

We also define the day- t BSI futures for stock index futures as:

BV futures ,t  SV futures ,t
BSI futures ,t 
BV futures ,t  SV futures ,t

Here, BV futures ,t is the buyer-initiated volume of stock index futures on day t , SV futures ,t is the seller-initiated

volume of stock index futures i on day t . A given day’s stock index futures BSI futures indicates whether most of

the investors are buyers ( BSI futures  0 , that is, the buyer-initiated volume is greater than the seller-initiated volume)

or sellers ( BSI futures  0 , that is, the seller-initiated volume is greater than the buyer-initiated volume).

The fourth proxy is based on psychological line index. Psychological line index is a sentiment indicator by Kim and

Ha (2010). Psychological line index is a good indicator for predicting short-term reversals of the futures market. Yang

and Gao (2014) also use psychological line index as stock index futures sentiment proxies.

The sample period ranges from April 16 2010 to April 26 2013. The data is from the China Financial Futures

Exchange (http://www.cffex.com.cn) and WIND database. The futures sentiment proxies are measured daily, weekly

and monthly.

This procedure leads to the stock index futures sentiment index ( S futures ), for example (IFL0, monthly):

S futrues ,t  0.253TVt  0.434OIt +0.338BSIt  0.316PSYt .

Each of the four components was standardized. The first principal component explains 78.19% of the (standardized)

sample variance, and only the first eigenvalue is far above 1.00, so we conclude that one factor captures the common

variation.

Table 1 Summary statistics.

Rspot R futures S spot S futures HML RMRF SMB
Mean -0.0334 -0.0351 -0.0713 -0.0355 -0.0002 -0.0003 0.0004
Median -0.0456 -0.0624 -0.1314 -0.0879 -0.0005 -0.0002 0.0012
Maximum 5.0490 5.7568 4.5455 3.5321 0.0238 0.0455 0.0203

8
 Minimum -6.3080 -7.8856 -2.0492 -3.4076 -0.0202 -0.0574 -0.0298
Std. Dev. 1.3685 1.3989 1.0281 1.1055 0.0053 0.0128 0.0059
Skewness -0.0796 -0.0012 0.5594 0.1855 0.4074 -0.2943 -1.0632
Kurtosis 4.8587 5.9841 3.3418 2.6506 4.4201 4.4871 5.9288
Jarque-Bera 147.61 377.70 58.05 11.02 113.71 108.50 555.63
Probability 0 0 0 0 0 0 0
Observations 734 734 734 734 734 734 734
The table 1 presents the summary statistics of the variables market returns, market sentiment, market behavior in
both spot market (CSI300 index) and futures market (CSI300 index futures). The sample period ranges from April 16
2010 to April 26 2013, covering a total of 734 days. Rspot denotes daily spot market returns. R futures denotes daily
futures market returns. Sspot denotes daily spot market sentiment. S futures denotes daily futures market sentiment.
RMRFt is the market return in excess of risk-free rate at time t , SMBt is the difference between the
value-weighted return of a portfolio of small stocks and the value-weighted return of a portfolio of large stocks at
time t , HMLt is the difference between the value-weighted return of a portfolio of high B/M stocks and the
value-weighted return of a portfolio of low B/M stocks at time t . We obtain daily returns on risk factors of
Shanghai Stock Exchange (SSE) which includes RMRF (market excess returns), SMB (small-minus-big firm
returns), and HML (high-minus-low returns) in RESSET database from April 16 2010 to April 26 2013.

Table 1 presents the summary statistics of the market returns, market sentiment in both spot market (CSI300 index)
and futures market (CSI300 index futures). The average of the daily futures market returns ( R futures ), the average of

the daily spot market returns ( Rspot ), the average of the daily futures market sentiment ( S futures ) and the average of the

daily spot market sentiment ( Sspot ) are below 0. It stands for the sample period is a bearish period.

9
 Fig. 1. The relationship between sentiment factors and returns in spot and futures market
(A) Panel A: Depicts the relationship between sentiment factors and returns in spot market (CSI300 index) from
April 16 2010 to April 26 2013, covering a total of 734 days. (B) Panel B: Depicts the relationship between
sentiment factors and returns in futures market (CSI300 index futures) from April 16 2010 to April 26 2013,
covering a total of 734 days.

3 Empirical results

3.1. The predictive power of mixed-frequency stock index futures sentiment

Following studies on sentiment-return predictability, such as those of Wang (2001), Simon and Wiggins III (2001)

and Wang (2003), we first consider simple univariate prediction models of stock index futures returns.

To test whether mixed-frequency stock index futures sentiment have more predictive power than stock

index futures sentiment on stock index futures returns, we organize our empirical work around the following

MIDAS regression model:

Rit*  i   FS ,i MS futures ,t 1   i ,t , (1)

where Rit* is the low frequency data on the left of equation is stock index future return①. It is necessary to

①
Rit* refers to the excess profit which removes both inflation and the week effect. The results are based upon the

10
mention that the horizon is the 1 order lag between period t  1 and period t (equal to one, for the monthly, weekly,

and daily regressions, respectively)②.

The null hypothesis is that  FS is zero. The alternative is that  FS is nonzero and reflects the predictive power of

mixed-frequency stock index futures sentiment. MS futures ,t 1  B( L1/ m ; )S futures

( m)
,t 1 , B( L
1/ m
; ) is a function with

parameter vector  , ( m)
L is the lag operator of high frequency data, S futures ,t 1 is the mixed-frequency futures

sentiment in future market, and m is the frequency multiple of mixed data. To introduce MIDAS regressions, suppose

that Rt is sampled at some fixed sampling frequency and call this the interval of reference. We have considered

some missing data and irregular data in our paper. For example, from April 4, 2011 to April 10, 2011, there are only

three trading data in this week due to Tomb-sweeping Day. So m is the frequency multiple of mixed data in the week,

m  3 . We construct weekly/daily mix-frequency sentiment in the week by the equation MS futures  B( L1/3 ; )S futures
(3)
.

In MIDAS regression model, B( L1/ m ; ) satisfies:

K
B( L1/ m ; )   B(k ; )Lk / m ,
k 0

where k is the lag order of High frequency data, and K is the maximum of k . We choose the exponent Almon

polynomial function:

exp(1k   2 k 2   pk p )
B .
 exp(1k   2 k 2   pk p )
K
k 1

The key point in MIDAS model is the choice of  and K , which will be showed in appendix A.

Table 2. The results of single market sentiment model.

Plane A

4
following model: Rit  R ft  ci  Week j  Rit* , i =1, , 4 Week j are the dummy variables checking for the week effect;
j 1

and Rit is the return of futures contract, R ft is the risk free rate.
②
We only choose 1 order lag. Because only short-term forecast power of sentiment is strong. Schwartz and Smith (2000)
and Mcclure et al. (2004) suggest irrational factors have more effect in short-term decision-making. Yang and Gao (2014)
find sentiment effect are both monotonous decreasing function of the time term on stock index futures returns.

11
 Frequency  FS t-stat Q-test 95% CI :  FS (  =1 ) ADJ. R2
t-test Q-test

Monthly 0.445 15.89 10.84 [0.36, 0.52] [0.39, 0.49] 0.198

Weekly 0.284 10.80 7.65 [0.20, 0.36] [0.22, 0.35] 0.102

Daily 0.240 8.53 5.41 [0.15, 0.32] [0.18, 0.33] 0.066

Plane B
Frequency  FS t-stat Q-test, 95% CI :  FS (  =1 ) ADJ. R2
t-test Q-test

Monthly/Weekly 0.497 19.60 15.48 [0.42, 0.57] [0.44, 0.58] 0.274

Monthly/Daily 0.500 20.33 16.45 [0.43, 0.57] [0.46, 0.57] 0.289

Weekly/Daily 0.447 14.62 12.21 [0.35, 0.53] [0.37,0.52] 0.173

Panel A of the table 2 presents the coefficients from the time-series regressions of the stock index futures market
return ( R futures ) as the dependent variable and the stock index futures sentiment as independent variable. Panel B of the
table 2 presents the coefficients from the single-factor time-series regressions of the stock index futures market return
( R futures ) as the dependent variable; the independent variables are mixed-frequency stock index futures sentiment.
We use monthly/weekly data, monthly/daily data and weekly/daily data in the mix-frequency regression. For
comparison, in the Panel B, we use monthly data, weekly data and daily data in the co-frequency regression. The
sample period ranges from April 16 2010 to April 26 2013, covering a total of 734 days, 154 weeks and 36 months.
The conventional t-statistics and Bonferroni Q-statistic of the coefficient by estimates are reported. Confidence
intervals that reject the null are in bold based on the t-test and Q-test at  =1 .

Table 2 presents the MIDAS regression model estimates at different frequencies, they are Monthly/Weekly,

Monthly/Daily and Weekly/Daily frequencies. Monthly/Weekly or Monthly/Daily denotes that the high frequency data

(weekly or daily) is mixed to the low frequency (monthly). Weekly/Daily denotes that the high frequency data (daily)

is mixed to the low frequency (weekly). All  FS coefficients are positive and statistically significant at different

frequencies. The persistence of the regressors is an important issue when analysing predictive regressions by

Campbell and Yogo (2006). So we choose the conventional t-test and the Bonferroni Q-test to inference predictability.

In the fifth column of the table, we report the 95% confidence intervals for  FS using the t-test. In the sixth column,

we report the 95% confidence interval for  FS using the Q-test at  =1 . The t-test and Q-test both reject the null of no

predictability for the stock index futures sentiment (  FS ) at all frequencies. So mixed-frequency stock index futures

sentiment effect is statistically significant on stock index futures returns.

12
 For test robust, we divide the sample into two equal subperiods, the first from Apirl 2010 to October 2011 and the

second from November 2011 to Apirl 2013. The results for the two subperiods are qualitatively similar to the results

for the entire sample period presented in Table 2. The coefficients of Monthly/Weekly regression model are 0.48 and

0.51 in the first and second subperiods, respectively, compared to 0.49 for the total sample period. The coefficients of

monthly frequency model are 0.43 and 0.45 in the first and second subperiods, respectively, compared to 0.44 for the

total sample period.

3.2. The predictive power of mixed-frequency stock index sentiment

Following Wang et al. (2011) and Yang and Gao (2014), they suggest that stock index futures sentiment and stock

index sentiment both affect stock index futures returns. We use MIDAS regression model to analyze whether

mixed-frequency stock index futures sentiment and mixed-frequency stock index sentiment have predictive

power on stock index futures returns ③:

Rit*  i  FS ,i MS futures ,t 1  SS ,i MSspot ,t 1   i ,t ， （t =1,…,T，i=1,...,4）, (2)

where  i represents a constant, and MSspot is the mixed-frequency stock index sentiment. MS futures  B( L1/ m ; )S futures
( m)
,

MSspot  B( L1/ m ; )Sspot

(m )
.

The null hypothesis is that the beta coefficients (  SS ,  FS ) of mixed-frequency stock index sentiment and

mixed-frequency stock index futures sentiment are zero, or any nonzero effect is due to rational compensation for

bearing systematic risk. The alternative is that the beta coefficients (  SS ,  FS ) are nonzero and reflects the predictive

power of mixed-frequency stock index sentiment and mixed-frequency stock index futures sentiment.

Table 3. The results of two market sentiment model.

Plane A
Frequency   FS  SS F-statistic ADJ. R2
Monthly -0.036 0.553 0.341 43.01 0.336
[-1.351] [20.106] [12.430]
Weekly -0.070 0.460 0.083 156.21 0.265

③
The serial correlation between the MS futures and MSspot is about 0.5. We consider MS futures and MSspot could reflect
futures market features and spot market features respectively.

13
 [-2.606] [16.689] [3.154]
Daily -0.003 0.305 0.251 10.64 0.133
[-0.101] [8.939] [6.818]
Plane B
Frequency   FS  SS F-statistic ADJ. R2
Monthly/Weekly 0.022 0.639 0.442 59.10 0.389
[0.801] [29.219] [20.207]
Monthly/Daily -0.055 0.766 0.550 36.84 0.492
[-2.388] [24.573] [18.999]
Weekly/Daily -0.013 0.491 0.360 15.57 0.304
[-0.445] [18.937] [15.661]
Panel A of the table 3 presents the coefficients from the time-series regressions of the stock index futures market
return ( R futures ) as the dependent variable; the independent variables are stock index futures sentiment and stock index

sentiment. Panel B of the table 3 presents the coefficients from the time-series regressions of the stock index futures
market return ( R futures ) as the dependent variable; the independent variables are mixed-frequency stock index futures

sentiment and mixed-frequency stock index sentiment. The t-statistics of the coefficient estimates are reported in
parentheses. We show the difference between  FS and  SS by F-statistic test. We present the F-statistic in

Appendix C. In the Panel A, we use monthly/weekly data, monthly/daily data and weekly/daily data in the
mix-frequency regression. For comparison, in the Panel B, we use monthly data, weekly data and daily data in the
co-frequency regression. The sample period of the market portfolio ranges from April 16 2010 to April 26 2013,
covering a total of 734 days, 154 weeks and 36 months.

Table 3 shows the results of multivariate prediction models at different frequencies. All  SS coefficients are

positive and statistically significant, which means that stock index sentiment has a significant influence to stock index

futures returns. Table 3 also presents the MIDAS regression model estimates at different frequencies in spillover

effect. All  SS coefficients (0.0914, 0.1089 and 0.0428) are positive and statistically significant at different

frequencies. So mixed-frequency stock index sentiment is also statistically significant on stock index futures returns.

In Table 3, the  FS coefficients are bigger than the  SS coefficient at all frequencies. The F-statistic test rejects

the null of no different predictability between  FS and  SS at all frequencies. It means that mixed-frequency stock

index futures sentiment is more important and significant on stock index futures returns than stock index sentiment.

3.3. The predictive power of mixed-frequency stock index sentiment in high and low sentiments periods

A large of empirical studies find consistent evidence that sentiment-driven investors participate and trade more

14
aggressively in high-sentiment periods (e.g. Yuan (2008), Karlsson, Loewenstein, and Seppi (2009), Yu and Yuan

(2011)). In this section, we provide further supports to the effect of mixed-frequency stock index futures sentiment

and mixed-frequency stock index sentiment in high and low sentiment periods. We use the model to analyze

whether mixed-frequency stock index futures sentiment and mixed-frequency stock index sentiment have

predictive power on stock index futures returns in different circumstances:

 ,i D1,t 1MS futures ,t 1  SS ,i MSspot ,t 1  SS ,i D2,t 1MSspot ,t 1   i ,t .

Rit*  i   FS ,i MS futures ,t 1   FS (3)

We define the following dummy variable to distinguish the optimistic sentiment and the pessimistic

sentiment:

0, MS futures ,t  0
D1,t   .
1, MS futures ,t  0

If D1,t  0 , the mixed-frequency stock index futures sentiment are pessimistic; if D1,t  1 , the

mixed-frequency stock index futures sentiment are optimistic.

0, MSspot ,t  0
D2,t   .
1, MSspot ,t  0

If D2,t  0 , the mixed-frequency stock index sentiment are pessimistic; if D2,t  1 , the mixed-frequency

stock index sentiment are optimistic.

For comparison, we use the model to analyze whether low-frequency stock index futures sentiment and

low-frequency stock index sentiment have predictive power on stock index futures returns:

 ,i D3,t 1S futures ,t 1  SS ,i Sspot ,t 1  SS

Rit*  i  FS ,i S futures ,t 1   FS  ,i D4,t 1Sspot ,t 1   i ,t . (4)

Where,

0, S futures ,t  0
D3,t   .
1, S futures ,t  0

If D3,t  0 , the stock index futures sentiment are pessimistic; if D3,t  1 , the stock index futures sentiment

are optimistic.

15
 0, Sspot ,t  0
D4,t   .
1, Sspot ,t  0

If D4,t  0 , the stock index futures sentiment are pessimistic; if D4,t  1 , the stock index futures sentiment

are optimistic.

Table 4. Regression results of stock index futures returns with heterogeneous sentiment.
Panel A Panel B
Variables 1 2 3 Variables 4 5 6
-0.125 -0.146 -0.297 -0.253 -0.123 -0.241
Intercept Intercept
[-4.328] [-2.759] [-5.736] [-5.012] [-3.392] [-4.963]
MS futures ,t 1
0.697 0.630 S futures ,t 1
0.400 0.452
[20.450] [29.007] [14.121] [16.564 ]
D1,t 1  MS futures ,t 1
0.232 0.166 D3,t 1  S futures ,t 1
0.253 0.082
[5.049] [4.524 ] [5.168] [3.153]
MSspot ,t 1
0.330 0.283 Sspot ,t 1
0.244 0.216
[19.776] [10.784] [8.682] [4.732]
D2,t 1  MSspot ,t 1
0.097 0.080 D4,t 1  Sspot ,t 1
0.109 0.052
[2.596] [1.750] [3.303] [1.899]
ADJ. R2 0.405 0.104 0.505 ADJ. R2 0.197 0.075 0.309
Panel A of the table 4 presents the coefficients from the time-series regressions of the stock index futures market
return ( R futures ) as the dependent variable; the independent variables are MS futures ,t 1 , D1,t 1  MS futures ,t 1 , MSspot ,t 1 and
D2,t 1  MSspot ,t 1 . Panel B of the table 4 presents the coefficients from the time-series regressions of the stock index
futures market return ( R futures ) as the dependent variable; the independent variables are S futures ,t 1 , D3,t 1  S futures ,t 1 ,
Sspot ,t 1 and . D4,t 1  Sspot ,t 1 . The t-statistics of the coefficient estimates are reported in parentheses. In the Panel A, we
use monthly/daily data in the mix- frequency regression. For comparison, in the Panel B, we use monthly data in
the co-frequency regression. The sample period of the market portfolio ranges from April 2010 to April 2013,
covering a total of 36 months.

Panel A of the table 4 presents the regression results of regressions of the stock index futures market return

( R futures ) on the mixed-frequency stock index futures sentiment MS futures ,t 1 , the effect of dummy variables on the

mixed-frequency stock index futures sentiment D1,t 1  MS futures ,t 1 , mixed-frequency stock index sentiment

MSspot ,t 1 , the effect of dummy variables on the mixed-frequency stock index sentiment D2,t 1  MSspot ,t 1 . The

D1,t 1  MS futures ,t 1 loading estimate is 0.283 with a t-statistic of 5.049 for the period of high mixed-frequency stock

index futures sentiment state. The D2,t 1  MSspot ,t 1 loading estimate is 0.097 with a t-statistic of 2.596 for the

period of high mixed-frequency stock index sentiment state. It shows the predictive power of mixed-frequency

sentiment is more significant in high sentiment periods.

16
 Panel B of the table 4 presents the regression results of regressions of the stock index futures market return

( R futures ) on the stock index futures sentiment S futures ,t 1 , the effect of dummy variables on the stock index futures

sentiment D3,t 1  S futures ,t 1 , stock index sentiment Sspot ,t 1 , the effect of dummy variables on the stock index

sentiment D4,t 1  Sspot ,t 1 . The D3,t 1  S futures ,t 1 loading estimate is 0.253 with a t-statistic of 5.168 for the period of

high stock index futures sentiment state. The D4,t 1  Sspot ,t 1 loading estimate is 0.109 with a t-statistic of 3.303 for

the period of high stock index sentiment state. It shows the predictive power of co-frequency sentiment is more

significant in high sentiment periods.

In this section, we find the predictive power of mixed-frequency and co-frequency sentiments are more

significant in high sentiment periods. The conclusion is consist with Yuan (2008), Karlsson, Loewenstein, and Seppi

(2009), Yu and Yuan (2011).

3.4. The predictive power of mixed-frequency stock index sentiment beyond Fama-French three factors

In this section, following Moskowitz et al. (2012), we add the past returns and the Fama-French three factors in

the model. And we examine whether the predictive power of mixed-frequency stock index futures sentiment

beyond the past returns and the Fama-French three factors. This consideration has been incorporated by allowing

the model in this study to include past stock index returns starting from 1 month to 4 months. The stock index

futures returns are regressed on the past stock index futures returns, mixed-frequency stock index futures

sentiment and the Fama-French three factors:

J
R futures ,t  i   FS ,i MS futures ,t 1   FR R futures ,t  k 1:t 1    j X j ,t 1   i ,t . (5)
j 1

Where R futures ,t k 1:t 1 defines the return for futures contracts from t-k-1 to t-1, which also is called time series

momentum factors. MS futures ,t 1 is the level of mixed-frequency stock index futures sentiment at time t-1 and

X j ,t 1 indicates other J control variables, such as the Fama-French three factors. A positive  FR indicates that there

exist persistent variations in the stock index futures price and that a momentum trading strategy will generate
positive profits. A positive  FS indicates that there exist persistent variations in the stock index futures price and that a
sentiment trading strategy will generate positive profits in stock index futures market.
For comparison, we use the model to analyze whether the past stock index returns; mixed-frequency stock

17
index sentiment and the Fama-French three factors have predictive power on stock index returns:

J
Rspot ,t  i   SS ,i MSspot ,t 1   SR Rspot ,t k 1:t 1    j X j ,t 1   i ,t . (6)
j 1

Where Rspot ,t k 1:t 1 defines the return for stock index from t-k-1 to t-1 and MSspot ,t 1 is the level of

mixed-frequency stock index sentiment at time t-1 and X j ,t 1 indicates other J control variables, such as the

Fama-French three factors.

Panel A of the table 5 presents the coefficients from the time-series regressions of spot market returns ( Rspot ) as

the dependent variable on mixed-frequency stock index sentiment. The control variables are the past stock index
returns and the Fama-French three factors. In specification (1), the adjusted-R2 between Rspot and Fama-French

three factors is only 0.152. Specifically, specifications (2), (3), (4) and (5) measure past returns by calculating the
return over the past 1st, 2nd,3rd and 4th months, respectively. All the results for specifications (2) to (5) are quite
similar. Past stock index returns and mixed-frequency stock index sentiment present a significantly positive
coefficient in these specifications, indicating that the better a stock’s past performance and the higher a investors’
past sentiment, the more returns will be. This finding also shows that investors in stock market intend to invest in
stocks with higher past returns, i.e., they adopt a momentum trading strategy.
Panel B of the table 5 presents the coefficients from the time-series regressions of futures market returns ( R futures )

as the dependent variable on mixed-frequency stock futures index sentiment. The control variables are the past
stock index futures returns and the Fama-French three factors. In specification (1), the adjusted-R2 between R futures

and Fama-French three factors is only 0.004, and the coefficient of Fama-French three factors is insignificant.
Specifically, specifications (2), (3), (4) and (5) measure past stock index futures returns by calculating the return over
the past 1st, 2nd,3rd and 4th months, respectively. All the results for specifications (2) to (5) are quite similar. Past
stock index futures returns and mixed-frequency stock index futures sentiment present a significantly positive
coefficient in these specifications, indicating that the better a stock index futures’ past performance and the higher a
futures investors’ past sentiment, the more futures returns will be. This finding also shows that investors in stock
index futures market intend to invest in stock index futures with higher past returns.
In the prior literature, Fama and French three factors have been a good correlation with contemporary asset returns.
In the paper, we focus on predictive power of variables, so we test whether Fama and French three factors have
predictive power in Chinese financial market. We find Fama and French three factors are just descriptive factors and
have smaller predictive power than sentiment factors. Mixed-frequency sentiment factors and time series
momentum factors have greater positively predictive power than Fama-French factors. This finding is common
to both spot and futures market.

Table 5. The predictive power of mixed-frequency sentiment factors, time series momentum factors and

18
Fama-French factors.
Panel
Panel B
A
Dep. Rspot ,t 1:t R futures ,t 1:t
Var.
1 2 3 4 5 1 2 3 4 5
0.248 0.318 0.357 0.382 0.327 0.340 0.377 0.397
MSspot ,t 1 [17.65 [19.89 [21.05 [21.84 MS futures ,t 1 [13.91 [14.49 [15.56 [16.11
7] 1] 8] 8] 7] 5] 3] 8]
0.269 0.245
Rspot ,t  2:t 1 [15.79 R futures ,t  2:t 1 [14.00
7] 0]
0.181 0.206
Rspot ,t 3:t 1 [13.10 R futures ,t 3:t 1 [15.42
4] 0]
0.132 0.146
Rspot ,t  4:t 1 [11.15 R futures ,t  4:t 1 [12.68
2] 2]
0.105 0.113
Rspot ,t 5:t 1 [10.01 R futrues ,t 5:t 1 [11.10
0] 8]
-10.32 -10.20 -10.24
26.242 -8.423 -1.854 -0.996 -2.409 -2.510 -1.513
3 0 2
HMLt 1 HMLt 1
[3.453 [-1.80 [-1.73 [-1.70 [-1.386 [-0.28 [-0.17 [-0.42 [-0.42 [-0.25
] 8] 0] 2] ] 5] 3] 5] 9] 4]
-10.93 -17.51 -15.74 -13.34 -10.92
53.294 -9.156 -7.691 -6.724 6.404
9 3 5 9 2
RMRFt 1 RMRFt 1
[17.49 [-4.41 [-3.59 [-2.96 [-2.573 [2.455 [-6.24 [-5.79 [-4.79 [-3.89
9] 2] 1] 7] ] ] 4] 9] 1] 1]
-21.11
7.178 5.779 2.927 3.169 -9.370 -0.552 1.262 -0.804 -1.807
4
SMBt 1 SMBt 1
[-2.98 [1.344 [1.049 [0.522 [-1.54 [-0.10 [0.238 [-0.14 [-0.32
[0.558]
3] ] ] ] 5] 2] ] 7] 5]
ADJ. ADJ.
0.152 0.336 0.292 0.263 0.247 0.004 0.221 0.247 0.198 0.171
R2 R2
Table 5 presents the coefficients from the time-series regressions of the stock index return ( Rspot ) in Panel A and
stock index futures return ( R futures ) in Panel B as the dependent variable. The explanatory variables are
mixed-frequency stock index futures sentiment, mixed-frequency stock index sentiment. The control variables are
the past returns and the Fama-French three factors. R futures ,t 1 k :t 1 defines the return for stock index futures from
t  1  k to t  1 . RMRFt is the market return in excess of risk-free rate at time t , SMBt is the difference
between the value-weighted return of a portfolio of small stocks and the value-weighted return of a portfolio of large
stocks at time t , HMLt is the difference between the value-weighted return of a portfolio of high B/M stocks and
the value-weighted return of a portfolio of low B/M stocks at time t . The t-statistics of the coefficient estimates are
reported in parentheses. We use monthly/daily data in the mix-frequency regression. All the intercept is
insignificant. The sample period of the market portfolio ranges from April 2010 to April 2013, covering a total of 36
months.

19
3.5. Out-of-sample return predictability
In in-of-sample test, we find mix-frequency stock index futures sentiment and mix-frequency stock index sentiment
both have predictability to stock index futures returns. These findings reflect data from the full sample and would not
have been available in real investing. To address this issue, it is necessary to test out-of-sample predictability measures.
We start our out-of-sample forecasts in June 2011, so the evaluation period is 2011:06-2013:04.
For each of the univariate mix-frequency sentiment prediction models, Table 6 reports out of-sample root mean
squared error (RMSE) values, measured relative to the RMSE obtained from the prevailing autoregressive benchmark
model. If values of univariate mix-frequency sentiment prediction models are smaller than the value of autoregressive
benchmark model, it suggests that a given model performs better than the benchmark. If values are bigger than
benchmark model, it suggests the opposite.
Table 6 presents the outcomes of RMSE of the mix-frequency sentiment prediction models. First, consider the
monthly prediction results in Panel A. For IFL0 as example, the RMSE for the first group of models including
monthly mix-frequency sentiment predictions MS futures (Monthly / Daily) , MSspot (Monthly / Daily) ,

MS futures (Monthly / Weekly) and MSspot (Monthly / Weekly) have the RMSE between 0.023 and 0.028, where the second

group of monthly frequency model generally has higher RMSE between 0.048 and 0.052. The benchmark models of
AR(1) method have the biggest RMSE 0.058. We find the RMSE of monthly mix-frequency sentiment predictions are
smallest in three groups, so the monthly mix-frequency sentiment predictions perform better than monthly frequency
sentiment predictions and benchmark model.
In Panel B, we also find the weekly mix-frequency model has smaller RMSE than weekly frequency model and
benchmark model. For IFL0 as example, the RMSE for the first group of models including weekly mix-frequency
sentiment predictions MS futures (Weekly / Daily) and MSspot (Weekly / Daily) have the RMSE between 0.011 and 0.012,

where the second group of weekly frequency model generally has higher RMSE between 0.024 and 0.025. The
benchmark models of AR(1) method have the biggest RMSE 0.026. We find the RMSE of weekly mix-frequency
sentiment predictions are smallest in three groups, so the weekly mix-frequency sentiment predictions perform better
than weekly frequency sentiment predictions and benchmark model. Compared with the in-sample evidence, the
out-of-sample results are robust that mixed-frequency sentiment factors have greater predictive power on stock index
futures market.

20
Table 6. Out-of-sample forecast performance: univariate sentiment prediction model
Plane A: Monthly
IFL0 IFL1 IFL2 IFL3
MS futures (Monthly / Daily) 0.023 0.022 0.023 0.021

MSspot (Monthly / Daily) 0.024 0.024 0.024 0.025

MS futures (Monthly / Weekly) 0.027 0.027 0.029 0.027

MSspot (Monthly / Weekly) 0.028 0.030 0.029 0.028

S futures (Monthly) 0.048 0.049 0.048 0.048

Sspot (Monthly) 0.052 0.053 0.053 0.051

AR(1) 0.058 0.058 0.058 0.058

Plane B: Weekly
IFL0 IFL1 IFL2 IFL3
MS futures (Weekly / Daily) 0.011 0.010 0.010 0.012

MSspot (Weekly / Daily) 0.012 0.012 0.012 0.013

S futures (Weekly) 0.024 0.025 0.025 0.024

Sspot (Weekly) 0.025 0.026 0.026 0.025

AR(1) 0.026 0.026 0.026 0.026

This table reports the ratio of the root mean squared forecast error of the mix-frequency sentiment prediction models
that include a constant and the predictor variable listed in each row, to the root mean squared error (RMSE) values of
the benchmark model. The forecast evaluation period is 2011:06-2013:04. The predictor variables include
monthly/daily mix-frequency stock index futures sentiment ( MS futures (Monthly / Daily) ), monthly/daily
mix-frequency stock index sentiment ( MSspot (Monthly / Daily) ), monthly/weekly mix-frequency stock index futures
sentiment ( MS futures (Monthly / Weekly) ), monthly/weekly mix-frequency stock index sentiment
( MSspot (Monthly / Weekly) ), monthly frequency stock index futures sentiment ( S futures (Monthly) ), monthly frequency
stock index sentiment ( Sspot (Monthly) ), weekly/daily mix-frequency stock index futures sentiment
( MS futures (Weekly / Daily) ), weekly/daily mix-frequency stock index sentiment ( MSspot (Weekly / Daily) ), weekly
frequency stock index futures sentiment ( S futures (Weekly) ), weekly frequency stock index sentiment ( S futures (Weekly) ).
Values smaller than benchmark model AR(1) indicate that the mix-frequency sentiment model performs better than the
benchmark.

21
4. The profitability of sentiment trading strategies

Moskowitz et al. (2012) find the performance of time series momentum trading strategy could provides a relatively

steady stream of positive returns that outperforms passive long positions. In section 3.4, we also show sentiment

factors and time series momentum factors have greater positively predictive power in in-sample tests. We next

investigate the profitability of trading strategies based on sentiment in out-sample forecast and compare the

profitability of (1) sentiment trading strategies, (2) momentum trading strategies and (3) passive long strategies.

(1) Sentiment trading strategies: we compute the time-t return based on the sign of the past cumulative sentiment

from time t-k-1 to t-1. For each contracts and day t, we consider whether the past cumulative sentiment over the

past k days is positive or negative. If the past cumulative sentiment over the past k days is positive, we go long

the contract. If the past cumulative sentiment over the past k days is negative, we go short the contract. We hold

the position for h days. For each sentiment trading strategy (k, h), we compute a single time series of daily

returns.

(2) Momentum trading strategies: we compute the time-t return based on the sign of the past cumulative return from

time t-k-1 to t-1. We derive this single time series of returns following the methodology used by Jegadeesh and

Titman (1993) and Moskowitz et al. (2012). However, different from computing monthly returns by Jegadeesh

and Titman (1993) and Moskowitz et al. (2012), we adjust the position everyday and compute the daily returns.

For each contracts and day t, we consider whether the cumulative excess return over the past k days is positive

or negative. If the past cumulative returns over the past k days is positive, we go long the contract. If the past

cumulative returns over the past k days is negative, we go short the contract. We hold the position for h days. For

each momentum trading strategy (k, h), we also compute a single time series of daily returns.

(3) Passive long strategies: we go long the contract, and hold the position for h days.

We vary both the number of days we turns to define the signal of the past cumulative return or the past

cumulative sentiment used to form the contract position (the ‘‘look-back period’’) and the number of days we hold

each futures contract after it has been formed (the ‘‘holding period’’). We then compute the time-t return based on

the sign of the past cumulative sentiment and the past return from t-k-2 to t-2, and so on until we compute the time-t

return based on the final past cumulative sentiment and the past return that is still being used from t-k-h to t-h. For

22
each (k, h), we get a single time series of daily returns by computing the average return. We focus on the properties

of the 30-day time series momentum strategy with 30-day holding period (e.g., k=30 and h=30),

Fig. 2 plots the Sharpe ratios to the profitability of sentiment trading strategies. For comparison, we also plot the

Sharpe ratios of a diversified momentum trading strategies and passive long strategies in all stock index futures

contracts. As Fig. 2 shows, the performance of sentiment strategy provides a relatively steady stream of positive

returns that outperforms momentum trading strategy and passive long positions. (Only one month cumulative return

of sentiment trading strategies is less than them of momentum strategy and passive long positions). Over the sample

July 2013 to June 2014, sentiment trading strategies has a statistically significant return, an annualized return of 80%

and annualized Sharpe ratio of 1.8, providing strong out-of-sample evidence of sentiment trading strategies.

Fig. 2. Sharpe ratios of passive long strategy, momentum strategy and sentiment strategy. Plotted are the monthly

Sharpe ratios of three strategy’s position in the stock index futures contract we study. Sample period is July 2013 to

June 2014.

23
5 Adding evidence of predict returns: the relation of sentiment and returns in futures market

We examine the ability of stock index futures sentiment and stock index sentiment to predict returns. Specifically, we

explore the bidirectional relation between intraday sentiment and near-term returns in a vector autoregression (VAR)

model by Brown and Cliff (2004). Brown and Cliff (2004) use VAR model to analyze the bidirectional relation between

sentiment and near-term returns at weekly and monthly frequencies in stock market. They reveal that market returns

clearly cause future changes in sentiment, while sentiment cloud not affect subsequent market returns. And they find the

relationships are consistent between monthly data and weekly data. However, we find the relationships are very different

between daily data. We show that futures returns clearly cause future changes in sentiment and sentiment also causes

future changes in futures returns in daily data. In short, sentiment and market returns have bidirectional relationships in

daily data.

5.1 Granger causality test

Granger causality test is a kind of econometric analysis tool to verify the causality between variables. We need

to confirm the causality between sentiment and returns in futures market before vector auto-regression (VAR).

Table 7. Granger causality test of sentiment and futures returns

This table reports the results of Granger causality test between sentiment and returns in futures market. The Granger
causality test is estimated using daily data. The sample period ranges from April 16 2010 to April 26 2013, covering a
total of 734 days.
Null Hypothesis F-Statistic Prob. N
1016
S futures does not Granger Cause R futures 100.53 0.00

R futures does not Granger Cause S futures 11.78 0.00

1016
S spot does not Granger Cause R futures 366.21 0.00

R futures does not Granger Cause S spot 10.85 0.00

Table 7 describes the results of Granger causality test between sentiment and market returns in futures. The

24
Ganger-causality test fails to reject the null hypothesis that S futures does not Granger Cause R futures at about 99%

level in first sets of row. The Ganger-causality test fails to reject the null hypothesis that R futures does not Granger

Cause S futures at about 99% level in second sets of row. The Ganger-causality test fails to reject the null hypothesis

that Sspot does not Granger Cause R futures at about 99% level in third sets of row. The Ganger-causality test fails to

reject the null hypothesis that R futures does not Granger Cause Sspot at about 99% level in fourth sets of row.

Overall, our tabulated results show that the sentiment index ( S futures , Sspot ) is the Granger Cause of market returns in

futures ( R futures ); market returns in futures ( R futures ) is also the Granger Cause of the sentiment index. In short,

sentiment and futures returns have bidirectional Granger Cause relationships in daily data.

5.2 The results of VAR model

Based on the results of Granger causality test between sentiment and returns in futures market, we use a

bivariate vector auto-regression (VAR) framework to explore how sentiment and the futures returns interact in

Chinese Market.

25
Table 8. Vector auto-regression (VAR) estimation of daily sentiment and the futures returns
This table reports the results from VAR estimation of the two indicated equations, where the dependent variable in
each specification is shown in the column heading. The VAR is estimated using daily data from April 16 2010 to April
26 2013, covering a total of 734 days.
R futures S futures R futures S spot

R futures ,t 1 0.212 -0.208 R futures ,t 1 0.035 -0.1294

[ 7.175] [-6.658] [ 1.185] [-5.758]
R futures ,t  2 0.239 -0.056 R futures ,t  2 0.165 -0.111
[ 8.773] [-1.964] [ 5.476] [-4.921]
S futures ,t 1 1.126 -0.709 Sspot ,t 1 0.613 -0.643
[ 37.531] [-22.345] [ 14.805] [-20.668]
S futures ,t  2 0.506 -0.209 Sspot ,t  2 0.537 -0.303
[ 11.741] [-4.590] [ 12.438] [-9.350]
Intercept -0.018 -0.003 Intercept -0.022 -0.007
[-0.671] [-0.112] [-0.582] [-0.240]
Adj.R 2 0.580 0.333 Adj.R 2 0.204 0.294

Table 8 reports the results from the estimation of the VAR system. Focusing first on the VAR system between

S futures indexes and R futures indexes, the results show that both the first and second lagged S futures indexes are

significantly positively related to the contemporaneous R futures index. Turning next to the second VAR system

between Sspot indexes and R futures indexes, we also see that the first lagged Sspot index is significantly positively

related to the contemporaneous R futures index. We find the ability of stock index futures sentiment and stock index

sentiment to predict returns in daily data.

6. Conclusions

Using the trading data in Chinese stock index futures market, based on the principal component analysis, this

paper constructs stock index futures sentiment and stock index sentiment at daily, weekly, and monthly frequencies.

Avoiding ignoring the information of high-frequency sentiment, we test the predictive power of mixed-frequency

26
stock index futures sentiment and mixed-frequency stock index futures sentiment on stock index futures returns at

different frequencies by using the MIDAS model.

The empirical results show that mixed-frequency stock index futures sentiment effect and mixed-frequency stock

index sentiment effect have positively predictive power statistically. By comparing mixed-frequency stock index

futures sentiment effect and mixed-frequency stock index sentiment effect, we find that mixed-frequency stock index

futures sentiment effect is more significant than mixed-frequency stock index sentiment effect. We also find

mixed-frequency stock index futures sentiment and mixed-frequency stock index sentiment has greater

positively predictive power in high sentiment period.

In out-sample forecasts, the mix-frequency sentiment predictions perform better than low-frequency sentiment

predictions and benchmark model by comparing RMSE. Compared with the in-sample evidence, the

out-of-sample results are robust that mixed-frequency sentiment factors have greater predictive power on stock

index futures market than low-frequency one. We also find the performance of sentiment strategy provides a

relatively steady stream of positive returns that outperforms momentum trading strategy and passive long positions.

The empirical results confirm that mixed-frequency stock index futures sentiment and mixed-frequency stock index

sentiment are systemic and important factors in futures price. Investors and the financial regulators could benefit from

understanding mixed-frequency sentiment in futures market. Our findings could raise some interesting issues for

future research. Future research may include forecast on futures prices with sentiment, macroeconomic and financial

predictors together.

Acknowledgements

This work was supported by the Natural Science Foundation of China (71471067), the Fundamental Research Funds

for the Central Universities (20148DXMPY03), the Doctoral Fund of Ministry of Education of China

(20120172110040), and the Chinese Ministry of Education and the Guangdong Province Project Fund

(XJ2JmN4100270). In addition, we wish to thank the reviewer for the valuable comments. All errors are the authors’

responsibility.

Appendix A. The choice of  and K

27
 Because of the result of MIDAS model would be influenced by the choice of  and K , so we choose optimal 

and K by  FS coefficient and adjust goodness of fit. The method how to select appropriate parameter is form

Ghysels et al (2004, 2007). In the Monthly/Weekly models, the  FS coefficient are best statistically significant and

adjust goodness of fit are biggest when we choose the lag order 4. We compare fast attenuation or slow attenuation. In

the Monthly/Weekly models, the  FS coefficient and adjust goodness of fit of fast attenuation are better statistically

significant than the  FS coefficient and adjust goodness of fit of slow attenuation. The results with fast attenuation

function are slightly better than the slow one, so the fast attenuation function is advisable.

Table 1. The results of the Monthly/Weekly MIDAS model with  and K

Lag order   FS ADJ. R2

K=4
-0.032 0.464*** 0.264
Slow
K=5
0.015 0.446*** 0.234
attenuation
K=6
0.032* 0.385*** 0.184
K=4
-0.034 0.497*** 0.274
Fast
K=5
0.026 0.458*** 0.248
attenuation
K=6
0.033* 0.457*** 0.2314
Table 1 presents the coefficients when we choose different  and K . The stock index futures market return ( R futures )
is the dependent variable; the independent variables are mix-frequency stock index futures sentiment. The sample
period ranges from April 16 2010 to April 26 2013, covering a total of 734 days, 154 weeks and 36 months.
*
denotes significance at 90% level,
**
denotes significance at 95% level,
***
denotes significance at 99% level.

Although the exponent Almon polynomial function is popular, there is some subjectivity. We change it to the

equally-weighted function. All the coefficients (  FS and  SS ) are positive and statistically significant at different

frequencies with the equally-weighted function. So the mixed-frequency stock index futures sentiment and

mixed-frequency stock index sentiment are also statistically significant with the equally-weighted function, which

means that the MIDAS model the equally-weighted function is also robust.

28
Table 2. The results of MIDAS model with the equally-weighted function

Frequency   FS  SS ADJ. R2

Monthly/Weekly 0.017 0.652*** 0.431*** 0.387

Monthly/Daily -0.036*** 0.717*** 0.527*** 0.454

Weekly/Daily -0.015 0.424*** 0.331*** 0.287

Table 2 presents the coefficients with the equally-weighted function in MIDAS model. The stock index futures market
return ( R futures ) is the dependent variable; the independent variables are mix-frequency stock index futures sentiment and
mix-frequency stock index sentiment. The sample period ranges from April 16 2010 to April 26 2013, covering a total of
734 days, 154 weeks and 36 months.
*
denotes significance at 90% level,
**
denotes significance at 95% level,
***
denotes significance at 99% level.

Appendix B. The stationary test

In order for this econometric analysis to make sense, we add the ADF tests to prove sentiment variables are

stationary. The variables include daily stock index futures returns ( R futures (Daily) ), weekly stock index futures returns

( R futures (Weekly) ), monthly stock index futures returns ( R futures (Monthly) ), monthly/daily mix-frequency stock index

futures sentiment ( MS futures (Monthly / Daily) ), monthly/daily mix-frequency stock index sentiment

( MSspot (Monthly / Daily) ), monthly/weekly mix-frequency stock index futures sentiment ( MS futures (Monthly / Weekly) ),

monthly/weekly mix-frequency stock index sentiment ( MSspot (Monthly / Weekly) ), weekly/daily mix-frequency stock

index futures sentiment ( MS futures (Weekly / Daily) ), weekly/daily mix-frequency stock index sentiment

( MSspot (Weekly / Daily) ). We find all of the variables are stationary at 1% level.

Table 3. Augmented Dickey-Fuller test statistic

Test critical values:
ADF test statistic 1% level 5% level 10% level
R futures (Daily) -33.419 -3.436 -2.864 -2.568
R futures (Weekly) -15.639 -3.458 -2.873 -2.573
R futures (Monthly) -8.026 -3.557 -2.916 -2.596
MS futures (Monthly / Daily) -15.907 -3.557 -2.916 -2.596
MSspot (Monthly / Daily) -7.409 -3.557 -2.916 -2.596
MS futures (Monthly / Weekly) -4.846 -3.557 -2.916 -2.596

29
 MSspot (Monthly / Weekly) -7.463 -3.557 -2.916 -2.596
MS futures (Weekly / Daily) -5.029 -3.458 -2.873 -2.573
MSspot (Weekly / Daily) -4.342 -3.458 -2.873 -2.573
This table reports the ratio of the Augmented Dickey-Fuller test statistic of the mix-frequency sentiment. The period is
2010:04-2013:04.

Appendix C. Testing multiple linear restrictions: The F test

We discuss the regression model with two variables. The estimated OLS equation is yt  b0  b1 x1,t 1  b2 x2,t 1   t .

The parameters b1 and b2 are the unknown coefficients. We say that two variables x1 and x2 has the positive

ability to predict stock index futures returns, if b1  0 and b2  0 .

We need to test if the parameters b1 and b2 are different statistically.

Our primary interest lies in testing the null hypothesis

H 0 : b1  b2  0 .

To determine a rule for rejecting H0, we need to decide on the relevant alternative hypothesis H1 : b1  b2  0

If H0 is true，we write the restricted model yt  b0  b1x3,t 1   ，where x3,t 1  x1,t 1 +x2,t 1 .

We compute the residual sum of squares in the restricted model： SSRH . 0

If H0 is not true，we write the unrestricted model yt  b0  b1 x1,t 1  b2 x2,t 1   t . We compute the residual sum of

squares in the unrestricted model SSRH . 1

( SSRH0 - SSRH1 ) / m
The F statistic is defined by F  F (m, n - p -1)
SSRH1 / (n - p -1)

We choose the degrees of freedom m  1 and p  2 .

References

Andersen, T. G., Bollerslev, T., & Lange, S. (1999). Forecasting financial market volatility: Sample frequency

vis-a-vis forecast horizon. Journal of Empirical Finance, 6(5), 457-477.

Andersen, T. G., Bollerslev, T., Diebold, F. X., & Labys, P. (2003). Modeling and forecasting realized volatility.

Econometrica, 71(2), 579–625.

30
Baker, M. & Stein, J. C. (2004). Market liquidity as a sentiment indicator, Journal of Financial Markets, 7, 271-299.

Baker, M. & Wurgler, J. (2006). Investor sentiment and the cross-section of stock returns, Journal of Finance, 61(4),

1645-1680.

Baker, M., Wurgler J., & Yuan, Y. (2012). Global, local and contagious investor sentiment. Journal of Financial

Economics, 104, 272-287.

Brown, G. W. & Cliff, M.T. (2005). Investor sentiment and asset valuation. Journal of Business, 78(2), 405-440.

Campbell, J. Y., & Yogo, M. (2006). Efficient tests of stock return predictability. Journal of financial economics, 81(1),

27-60.

Cen, L., Lu, H., & Yang, L. (2013). Investor Sentiment, Disagreement, and the Breadth-Return Relationship.

Management Science, 59(5), 1076-1091.

Fong, W. M., & Toh, B. (2014). Investor Sentiment and the MAX Effect. Journal of Banking & Finance. 46, 190-201.

Ghysels, E., Santa-Clara, P. & Valkanov, R., (2004). The MIDAS Touch: Mixed Data Sampling Regressions, Mimeo,

Chapel Hill, N.C.

Ghysels, E., Sinko, A. & Valkanov, R. (2007). MIDAS Regressions: Further Results and New Directions, Econometric

Reviews, 26, 53-90.

Greenwood, R., & Shleifer, A. (2014). Expectations of returns and expected returns. Review of Financial Studies,

27(3), 714-746.

Hallam, M., & Olmo, J. (2014a). Forecasting daily return densities from intraday data: A multifractal approach.

International Journal of Forecasting, 30(4), 863-881

Hallam, M., & Olmo, J. (2014b). Semiparametric density forecasts of daily financial returns from intraday data.

Journal of Financial Econometrics, 12(2), 408–432.

Jegadeesh, N., & Titman, S., 1993. Returns to buying winners and selling losers: implications for stock market

efficiency. Journal of Finance 48, 65–91.

Joseph, K., Babajide Wintoki, M., & Zhang, Z. (2011). Forecasting abnormal stock returns and trading volume using

investor sentiment: Evidence from online search. International Journal of Forecasting, 27(4), 1116-1127.

Kim, J. S., Ryu, D., & Seo, S. W. (2014). Investor sentiment and return predictability of disagreement. Journal of

Banking & Finance, 42, 166-178.

31
Kim,T. & Ha. N. (2010). Investor sentiment and market anomalies, SSRN, Working Paper.

Kumar, A. & Lee, C. (2006). Retail investor sentiment and return comovements, Journal of Finance, 61, 2451-2486.

Kurov, A. (2010). Investor sentiment, the stock market’s reaction to monetary policy. Journal of Banking Finance,

34(1), 139-149.

Lewellen, J., 2004. Predicting returns with financial ratios. Journal of Financial Economics 74, 209–235.

Li, J. (2014). Multi-period sentiment asset pricing model with information.International Review of Economics &

Finance, 34(C), 118-130.

Moskowitz, T. J., Ooi, Y. H., & Pedersen, L. H. (2012). Time series momentum. Journal of Financial

Economics, 104(2), 228-250.

Safa, M. F. & Maroney, N. C. (2012). Bid-Ask Spread, Futures Market Sentiment and Exchange Rate Returns. Journal

of Economic Cooperation and Development, 33, 63-85.

Scheinkman, J. & Wei, X. (2003). Overconfidence and speculative bubbles, Journal of Political Economy, 111,

1183-1219.

Schmeling, M. (2009). Investor sentiment and stock returns: some international evidence. Journal of Empirical

Finance, 16 (3), 394-408.

Simons, D. P., & Wiggins, III R. A. (2001). SP futures returns and contrary sentiment indicators, Journal of Futures

Markets, 21, 447-462.

Stambaugh, R.F., 1999. Predictive regressions. Journal of Financial Economics 54, 375–421.

Verma, R. & Soydemir, G. (2009).The impact of individual and institutional investor sentiment on the market price of

risk. The Quarterly Review of Economics and Finance, 49(3), 1129-1145.

Wang, C. (2001). Investor sentiment and return predictability in agricultural futures markets, Journal of Futures

Markets, 21, 929-952.

Wang, C. (2003). Investor sentiment, market timing, and futures returns, Applied Financial Economics, 13, 891-898.

Wang, Y. H., Keswani, A. and Taylor, S, J. (2006). The relationships between sentiment, returns and volatility,

International Journal of Forecasting, 22(1), 109-123.

Wang, Y. M., Li, C. A., & Lin, C. F. (2011). Investor sentiment spillover effects between the futures and spot markets,

Investment Management and Financial Innovations, 8, 63-71.

32
Yang, C. & Gao, B. (2014). The term structure of sentiment effect in stock index futures market. The North American

Journal of Economics and Finance, 30, 171-182.

Yang, C. & Zhang, R. (2014). Does mixed-frequency investor sentiment impact stock returns? Based on the empirical

study of MIDAS regression model, Applied Economics, 46, 966-972.

Yang, C., & Zhou, L. (2015). Investor trading behavior, investor sentiment and asset prices. The North American

Journal of Economics and Finance, 34, 42-62.

Yu, J. & Yuan, Y. (2011). Investor sentiment and the mean-variance relation. Journal of Financial Economics, 100(2),

367-381.

33