关键词: 投资者情绪, 时间序列动量, 主成分分析, 偏最小二乘

Abstract: Momentum effect is one of the most typical market anomalies. The effectiveness of cross-sectional momentum strategies and time series momentum strategies has been verified in many studies. With the development of behavioral finance, the significant impact of investor sentiment on asset pricing and momentum effect has been pointed out in recent studies. Some researchers have improved the cross-sectional momentum strategies and the mean reversion strategies by incorporating investor sentiment, but no studies have attempted to introduce investor sentiment into the construction of time series momentum strategies. Therefore, this paper proposes to construct sentiment-based time series momentum strategies, so as to provide investors with potentially profitable investment opportunities.
As for the benchmark time series momentum (TSM) strategy, we use the arithmetic averageof historical returns to generate trading signals. Specifically, for each instrument i and month t, we consider whether the excess return over the past J months is positive or negative and go long if positive and short if negative at the end of the month, holding the position for K months. Here J is referred to as the length of the formation period, and K is referred to as the length of the holding period. For each trading strategy with time parameter (J,K), the strategy return during month t is the average return across all “active” portfolios, namely the average return on the portfolios that were constructed last month, the month before (if still held in month t), and so on.
While constructing our sentiment-based time series momentum (STSM) strategy, we use the Sharp Ratio difference of the benchmark TSM strategy during the high sentiment period and the low sentiment period (ΔSR) to adjust the position longed and shorted in the benchmark TSM strategy. Specifically, in the high sentiment period, if the TSM trading signal is “long”, we should increase the long position by ΔSR; if the TSM trading signal is “short”, we should reduce the short position by ΔSR. In the low sentiment period, if the TSM trading signal is “long”, we should reduce the long position by ΔSR; if the TSM trading signal is “short”, we should increase the short position by ΔSR. The judgment of high and low sentiment periods is based on the lower and higher 30% percentiles.
The empirical study uses the monthly data of the constituent stocks of the small and medium enterprise (SME) 100 index. After excluding one stock with insufficient data, 99 stocks are employed. The sample period ranges from January 2010 to July 2021, altogether 139 trading months. We form a composite sentiment index from seven market indices, i.e., discount rate of closed-end funds, market trading volume, amount of IPOs, first day earning, number of newly opened accounts, consumer confidence index and market turnover rate, applying the principle component analysis method. Considering the fact that some indices take longer to reveal the same sentiment, we start by estimating the first principal component of the seven indices and their one-month lags. This gives us the first-stage composite sentiment index C14. We then compute the correlation between the first-stage index and the current and lagged values of each of the seven indices, and construct the composite sentiment index C7 from the seven market index variables with higher correlation with the first-stage index, controlling macro-economic effects. The first four principal components are selected so as to explain at least 85% of the variance. Time series plots confirm that the composite sentiment index has similar trend to that of the SME index and leads the SME index to some extent. Applying the VAR model for the composite sentiment index and the SME index reveals that the composite sentiment index has significantly positive impact on future SME index return, supporting our design of the STSM strategy.
We plot cumulative excess returns for sixty-four common (J,K) combinations, that is, J=1, 3, 6, 9, 12, 24, 36, 48 months, K=1, 3, 6, 9, 12, 24, 36, 48 months, and chose the six (J,K) combinations with superior performance, i.e., (6,3), (6,6), (9,1), (9,3), (9,6) and (12,1), as benchmarks to construct our STSM strategies. Comparing the cumulative excess returns of the STSM strategies with those of the TSM strategies suggests that, the sentiment-based STSM strategies do have much better performance. The hypothesis test results further confirm that the performance gains of the STSM strategies are all significant, and the STSM strategies also have significantly higher return than the buy-and-hold strategy. As for the robustness test, we use different lengths of the investment horizon, the partial least squares method instead of the principal component analysis method for sentiment index synthetization, and the weighted average returns instead of the arithmetic average returns for generating trading signals, respectively, and all get consistent results.
This study successfully constructs sentiment-based time series momentum strategies, which not only provides tools for investors' asset allocation, but also sheds lights on future quantitative studies. Current composite sentiment index is constructed from market indices, which indirectly reflects all market participants' sentiment. Recent researches suggest that using text data such as stock forum comments and analyst reports can construct indicators that directly reflect the subjective sentiments of different types of investors. Therefore, our future work will explore the application of richer investor sentiment indicators in time series momentum strategies. In addition, we will also explore the effective introduction of richer investor sentiment indicators in commonly used quantitative investment strategies such as cross-sectional momentum strategies and reversal strategies, in order to provide investors with more effective asset allocation tools.

Key words: investor sentiment, time series momentum, principal component analysis, partial least squares