การทำนายทิศทางของราคาฟิวเจอร์สของดัชนี SET50 ในตลาดสัญญาซื้อขายล่วงหน้า (ประเทศไทย) ด้วยแบบจำลองอะแด็พทีฟโลจิต

Abstract

This study aims at the analysis of determinants of price direction of SET50 Index Futures (TFEX) in the Thailand Futures Exchange using Logit model, prediction the direction using both Logit and Adaptive Logit models, and comparison of the prediction accuracy of both models. Major findings are as follows: 1) The most influential determinants of the price direction of TFEX are the change of SET index of the previous day which positively affects the price direction, and the level of SET50 index of the previous day which negatively affects the price direction. 2) Logit model yields the highest accuracy rate at around 47.14 per cent when tested by the data in the validation set. 3) Adaptive Logit model also yields the 47,14 percent of accuracy rate when assigning the learning rate at 1. However, the adjustment of higher learning rate to be 100 drives the model to yield more accuracy rate when tested by the data in the validation set. 4) The comparison of accuracy rates between both models reveals that they yield equal accuracy rates at 57 per cent when tested by the data in the testing set. However, the prediction is extreme such that both models predict just only the rise of the price for every day. 5) Other lessons learnt that bring about better understanding of technical matters of Logit and Adaptive Logit models such as, first, the selection of just significant variables from Logit model for the prediction yields better performance than the selection of all variables. Second, Adaptive Logit model does not produce the overfitting condition. Third, although some cases of Adaptive Logit model do not converge, the solution for this problem is to add the constant into the model. Last, Adaptive Logit model can turn the poorly performed model in terms of accuracy of the prediction into the highly perform model. This is because the algorithm of the Adaptive Logit model that tries to perfection the parameters of each observation to meet the prediction and the true value. The more accuracy may be achievable by classification of observations whose explanatory variables are similar, and then apply the specific parameters to each group rather than using the average values of the parameters. This method is a challenge for future study that may lift the performance of Adaptive Logit model over the Logit model.