An Empirical Research For Factor Model Based On Intraday Market Data

In the practice of investment,due to the development of risk factor model,the concepts of Beta and Alpha have been well known.Beta is part of asset returns that can be explained by one or more risk factors,while the rest is called Alpha.Beta has more and more eaten into the whole part of asset returns since new risk factors had been discovered and as a result,Alpha is being dissipating.From the academic perspective,new risk factors assist people to understand the source of asset returns,while helping achieve excess returns from a practical view.Another motivation for the thesis is the market.As is known,in the first half year of 2017,China stock market had a structural differentiation.A Chinese version of "Nifty Fifty" including white spirit companies?home electronics companies was emerging,while relatively small stocks which had been in fast growth were now declining.Nearly eighty percent Public Offering of Quantitative Funds had performance deficits,most of which were using stock selection strategy based upon multi-factor models.It follows that the traditional factor system as a whole has high correlation with the size factor.In other words,the number of factors which have independently contributions is far from enough.Once the market style transforms,the whole system is exposed to huge risks.Hence,it is of high practical value to mine new factors that are less correlated with the traditional factor zoo in a systematic and effective way.This thesis uses a brand new factor system formally called Transactional Alpha system and high-frequency market data are used for constructing factors.This type of factor system was raised by Igor Tulchinsky et al.in 2015.In this system,a factor is a prediction model of the stock return for the next trading day based on historical data.The absolute value of the factor is directly associated with the position,the bigger the value,the heavier the position,a positive sign means longing the stock,while a negative sign shorting the stock.Then a formal backtesting is carried out to address the effectiveness of this particular factor.A distinctive character of this system is that factors are mathematical expressions,hence it's also called Formulaic Alpha system.In the domestic,this type of factor system is published for the first time by a financial engineering research group from GUOTAI JUNAN SECURITIES.They replaced the old "Formulaic Alpha System" by calling it"Transactional Alpha System".On the other hand,high frequency trading(HFT)is limited,due to regulations and risk control requirements.Stepping back,the method of using high frequency data to extract low-frequency signals is getting popularity,yet published research papers related to this new factor system only encompassed the low-frequency data,namely daily "OHLCV" data.This thesis incorporates the method above into Transactional Alpha System.Under this system,we construct two new types of factors from intraday data.Specifically,the first type of factors is moneyflow indicators constructed from tick data and the second type of factors are relative trading price indicator reflecting institutions' sentiments about the market.These two factors try to capture the investors' behavioral biases.Since intraday data reflect the market micro structure more closely,while interday data lose almost all information on this aspect,intraday data have richer implications for predicting assets' returns than interday data do.We carry out a formal backtesting and significance tests showing that these new factors have relatively stable and high risk premium and beyond that have low correlation(around 0.2)with the size factor.This is a huge complement to the traditional factor model.Moreover,these factors also reveal that speculations have their place in Chinese stock market.Institutional and individual investors have preference for short-term trading which has some fixed pattern.The factors constructed in this thesis exploit this kind of behavioral bias for excess returns.It should be mentioned that any seemingly solid patterns have possibility of altering.In other words,no matter how effective the factor was in history,it could be invalid in the future.Therefore,we further assess the factor's effective period by looking at the degree of decay of out-of-sample performance and by comparing the predictive power of factors with different lags.The results illustrate that the predictive power of factors constructed from intraday data decays as lags increase.The effective period of predicting future returns based on these factors lasts at most five to six trading days,i.e.factors have no predictive power for stock returns beyond six trading days.Besides,by looking at monthly information ratio,the relative-price type factor suffers from the declining of out-of-sample performance with respect to in-sample performance,while moneyflow type factors do not show a downward trend in terms of monthly information ratio significantly,which implies that these factors have longer effective periods and are expected to be more satisfactory in the future.