410 likes | 501 Views
predictability of non-linear trading rules in the us stock market chong & Lam 2010. Overview. Liu Min Qi Yichen Zhang Fengtian. Literature review of the paper A brief introduction of the 3 models used in the paper The strategies and results Model selection for our project
E N D
predictability of non-linear trading rules in the us stock marketchong & Lam 2010
Overview Liu Min QiYichen Zhang Fengtian • Literature review of the paper • A brief introduction of the 3 models used in the paper • The strategies and results • Model selection for our project • Data selection for our project • Statistical tests for data selected • Model regression techniques (parameter estimation and trading strategies for SETAR model) • Algorithm in java • Demonstrate SETAR model (JAVA program) • Results and analysis
Literature Review • The paper examines the effect of 3 models • Namely • Variable Moving Average (VMA Model) • Autoregressive Model (AR Model) • Self-Exciting Threshold Autoregressive Model (SETAR Model) • On four major US indices • Dow Jones Industrial Average (DJIA) • NASDAQ composite index • New York Stock Exchange composite index • Standard and Poor 500 index (S&P 500)
VMA Model • General Form • VMA(S, L) • Where S represents the short term window • And L represents the long term window • Calculated from the below formula • for a n day moving average • Models used in the paper includes VMA(1, 50), VMA(1, 150) and VMA(1, 200)
AR Model • General Form • Commonly referred as AR (p) Model • Where p is the order of autoregressive part
AR Model • A linear time series model • For the paper, an AR (1) Model was used • ∆Yt is the natural log difference of the stock index • Where ∆Yt= Yt- Yt-1 • Alphas are the fitted coefficient • εt is the residual error • AR (1) was chosen because the estimated coefficients are significant, suggesting it is good enough for modelling dynamics of return series
SETAR Model • General Form: • Usually referred as SETAR(k, p) model • k is the number of regimes • p is the order of the autoregressive part
SETAR Model • A non-linear time series model • An extension of AR models • Higher degree of flexibility due to the threshold parameter • Which introduces a regime switching behavior
SETAR Model • For the paper, a SETAR(2, 1) model was chosen • ∆Yt is the natural log difference of the stock index • Where ∆Yt= Yt- Yt-1 • Alphas and betas are the fitted coefficient • εt is the residual error • d is the delay factor • γ is the threshold parameter
SETAR Model • Why SETAR (2, 1)? • Because (as claimed by paper) • It is simple and has good predictability • Threshold parameter already captures non-linearity • Therefore additional benefit of higher autoregressive order is small • The estimated coefficients are significant based on statistical tests • Suggesting first order model is good enough to describe dynamics of the return series
Strategies • For VMA • Pretty straightforward • Buy if MA(S) > MA(L) • Sell if MA(S) < MA(L) • For AR and SETAR • Model fitting is required for every w observations • Buy if > 0 • Sell if < 0
Dow Jones Industrial Average index. ‘Buy>0’ and ‘Sell>0’ are the fraction of positive buy and sell returns. Buy , Sell and Buy-Sell columns show the one day conditional mean for buy, sell and buy-sell returns
Results • Performed using observation window period of: • 50, 150, and 200 days • SETAR performed slightly better than AR for DJIA and S&P 500 • AR performed slightly better in NASDAQ • Both SETAR and AR outperformed VMA
Model Selection • Therefore, SETAR model was chosen for our project • Because of the better results obtained from the paper • And also because of its non-linearity • Which gives it flexibility in modelling • However, simulation may be slow due to a need for multi-parameter fitting for each signal
Data Selection • We had chosen the HK’s Hang Seng Index and Singapore’s Straits Times Index • Data selection (from yahoo finance) • Hang Seng Index • Daily closing price from 31st Dec 1986 to 31st Dec 2010 • Total 5962 Observations • Straits Times Index • Daily closing price from 31st Dec 1987 to 31st Dec 2010 • Total 5754 Observations
Index Statistics • Summary statistics for daily log returns – full sample • JB stat represents the Jarque-Bera test for normality • ρ(i) is the estimated autocorrelation at lag i • Q(5) is the Ljung-Box Q statistic at lag 5 • Numbers marked with * are significant at 1% level
Statistical Results • From the values of skewness, kurtosis, and Jarque-Bera statistics • Returns are leptokurtic, skewed, and not normally distributed • Ljung-Box Q statistics at 5th lag significant at 1% • Suggestive of substantial serial correlation in stock returns • Essential for existence of trading-rule profits • These results are consistent with that found in the main paper • Which may be indicative of the model’s efficiency on the Hang Seng Index and Straits Times Index
Parameter estimation • Model:
Parameter estimation • Use Ordinary Least Square method to find γ and θ. (Refer to Bruce E. Hansen (1997) Inference in TAR Models. Studies in Nonlinear Dynamics & Econometrics, Volume 2, Issue 1)
Parameter estimation Remarks: In our case, d (delay parameter) = 1. Observe that the residual variance only takes on at most ndistinct values as γis varied, we set γ = ΔYt-d, t = 2,…,n.
Parameter estimation • Thus, the estimate of θ is Given n observations, we use OLS to obtain the fitted coefficients γ and θ andpredict ∆Yt+1 based on ∆Yt.
Trading strategy • The SETAR trading strategy is as follows: where W is the observation window and is the conditional expectation of ΔYt+1 based on most recent W observations up to day t.
Trading strategy • Remarks: • Just imagine that we use the model to predict the price tomorrow. If the predicted price is higher than today actual price, then we buy. Otherwise, we sell. • The value of αand βchange with the observation window, as we use the most recent w observations. So as we move, we roll the window forward and update the α and β to get the next prediction of ΔY.
Trading strategy • For example, given W = 50 and n = 100, • 1. Obtain γ and α0 α1 β0β1. • 2. Obtain ΔYt+1 based on ΔYt and estimated parameters. • 3. Buy if ΔYt+1 > 0. Sell if ΔYt+1 < 0. • 4. Shift the observation window (set t = t+1) and repeat Step 1 to Step 3.
Algorithm in java • The model
SETAR(1, 50) model Predicted stock index moving by time - HANG SENG INDEX
SETAR(1, 50) model Predicted stock index moving by time - STRAITS TIMES INDEX
SETAR(1, 150) model Predicted stock index moving by time - HANG SENG INDEX
SETAR(1, 200) model Predicted stock index moving by time - HANG SENG INDEX
Empirical results of implementing the trading strategies on the HANG SENG INDEX
Empirical results of implementing the trading strategies on the STRAITS TIMES INDEX
Future work T-Statistics • Mean return of buy periods. • Mean return of sell periods. • Buy- sell return. AR Model • The performance of the nonlinear trading rule (SETAR) is compared with that of the linear model (AR).