基於反向情緒指標使用分類元系統預測美國S&P期貨指數

Abstract

由於近年來金融市場與經濟環境的快速變動與成長，進行投資決策已成為一件非常困難的事。許多文獻使用迴歸模型或ARCH及GARCH模型等傳統統計方式來預測股票或期貨的價格走勢，但這些統計模型常以許多不合理的假設作為前提。而最近數十年蓬勃發展的人工智慧技術不但需要較少不合理的假設且其運作方式更接近人類思維模式。在所有人工智慧技術中，學習分類元系統整合了基因演算法及強化式學習，可在規則集中區別出適合的規則並發掘更新及更好的規則。學習分類元系統的動態及演化之特性使其非常適合用於預測具有高度不確定性的財務時間序列。因此，本研究使用新一代的學習分類元系統(延伸分類元系統)建立一個期貨投資交易模型。 雖然許多研究學者都利用以技術指標作為輸入變數的人工智慧模型尋找價格趨勢，但只有極少數研究著重於以情緒指標為輸入變數的人工智慧技術之預測效能。先前的研究顯示volatility index、put-call ratio及trading index等反向情緒指標具有極佳的預測能力。因此，本研究利用基於反向情緒指標的延伸分類元系統預測美國S&P 500期貨指數，希望能挖掘出財務預測領域中的深層知識。 為了驗證以情緒指標為輸入變數的延伸分類元系統是否適合用來預測財務時間序列，本研究使用樣本外迴歸模型及標竿模型做為實驗對照組，標竿模型包含買進即持有、趨勢追隨及均數反轉等交易策略。為了使實驗環境更接近真實交易環境，在所有模擬交易中都考慮了交易佣金與買賣價差。實驗及統計檢定結果顯示基於反向情緒指標的延伸分類元系統在與其他比較模型相較之下具有極佳的預測準確度及獲利能力。

Nowadays, the complexity of economic markets has grown rapidly, so making the investment decision has become a very difficult task. Previous studies of financial time series forecasting utilized statistical approaches such as regression model, ARCH or GARCH model to predict stock or futures price. Compared to these statistical models, artificial intelligence (AI) techniques have less unrealistic assumptions and operate closer to the way of human thought. Among the AI techniques, learning classifier system (LCS) combines genetic algorithm and reinforcement learning to distinguish fit rules from the rule set and extract newer and better rules. The dynamic and evolutionary property of LCS could fit the highly uncertain environment such as financial time series. Accordingly, this thesis applies extended classifier system (XCS), the newer version of LCS, to construct a futures trading model. Although many researchers have applied AI models utilizing technical indicators as input variables to discover the price trend, relatively little research has focused on the forecasting performance of AI techniques based on sentiment indicators. The previous research showed the contrary sentiment indicators, including volatility index, put-call ratio, and trading index, have excellent predictability. Therefore, this thesis utilizes XCS to model the contrary sentiment indicators, attempting to dig out the in-depth knowledge of the financial forecasting domain. To prove that XCS based on contrary sentiment indicators could apply to forecast financial time series, the performance of XCS is compared against that of out-of-sample regression model and benchmark models, containing buy-and-hold, trend-following, and mean-reversion trading strategies. In the experiments, commissions and bid-ask spreads are accounted for on all transactions to make the experiments closer to the real transaction. The simulation and statistical results showed that XCS based on contrary sentiment indicators possesses more excellent forecasting accuracy and profits earning capability than other comparison models.