建構演化式類神經網路於財務時間序列信賴區間動態風險值評價模型

Abstract

財務時間序列的資產價格與報酬率波動之評價一直是財務經濟學者所關心的主題。自從 Engle(1982)提出自我迴歸條件異質變異模型(ARCH模型)及 Bollserlev(1986)提出之一般化自我迴歸條件異質變異模型(GARCH模型)之後，ARCH及其衍生模型即被廣泛的運用在財務時間序列資料的分析上，且證明資產報酬率之波動有隨時間改變而改變的特性。除此之外，有些文獻亦指出金融市場具有動能的物理現象，然而，GARCH模型僅以條件變異數與殘差平方項做為解釋波動性的變數，而未將其他會影響到波動性的行為財務因素考量在內。近年來，人工智慧於金融市場之應用，不同於傳統財務數學與統計模型存在許多的假設前題與限制，相當適合於求解複雜的非線性問題。本研究提出一適用於財務時間序列資料之區間式類神經網路，除改良傳統類神經網路之點估計缺點，並保留類神經網路之非線性預測特性，補強GARCH僅以變異數與殘差項為自變數的不足，加入財務時間序列的價量技術指標與變異數之動能物理現象作為解釋變數，藉此加強資產價格/報酬波動性預測模型之可解釋性。 另外，國際間金融自由化的趨勢、法規的鬆綁與金融環境的開放，亦使得金融市場之風險管理議題日趨重視，而風險值(Value at Risk)具有量化風險的優點，而且廣為投資人及金融機構所採用。故本研究以區間式類神經網路為基礎，建構動態式風險值評價模型，除了強化捕捉金融資產價格/報酬波動的叢群現象之外，並進一步提供投資人與金融機構作風險控管之用。本研究以台灣第一個ETF商品—台灣50ETF為實證資料，經實證結果與傳統GARCH模型相較發現，以MSE、MAE與MAPE為衡量標準，預測績效最佳為區間式類神經網路模型。除此之外，以準確性、保守性與效率性三者比較，區間式類神經網路模型具有較優的準確性，GARCH模型具有較佳的保守性與效率性，但在沒有兼具準確性的前提下，GARCH模型有較佳的保守性與效率性似乎是沒有意義的。 總而言之，本研究提出創新的區間式類神經網路架構，來進行資產/價格報酬的動態風險值之評價，能有效應用於財務時間序列資料之最大可能發生損失之風險控管，其中採用有意義的財務時間序列的價量技術指標與變異數之動能物理現象，增強時間序列預測模型的可解釋性，以做為投資人與金融機構於投資決策輔助之用。此方法未來可加入更多解釋指標，並廣泛應用於其他領域以提升良好的風險管理機制。

The appraisement of asset price/return and volatility has always been the concerned topic in field of financial time series by the financial economists. After the introduction of the Autoregressive Conditional Heteroscedastic Model (ARCH) by Engle (1982) and the extension version of Generalized Autoregressive Conditional Heteroscedastic Model (GARCH) by Bollserlev (1986), these two models are wildly adopted in analyzing financial time-series data, and further prove the feature of “time-varying” of return on assets. In addition, several literatures also pointed out the physical phenomenon, “momentum”, exists in financial market. Nevertheless, the GARCH model only considers conditional variance and sum of squared residual as the variables in explaining volatility, but fails to cover the influence of factors of behavioral finance on volatility. In recent years, the applications of artificial intelligence in financial market are appropriate for solving complicated non-linear problems, since they do not have to deal with many assumptions and limitations within conventional financial mathematics and statistical model. This research introduces an interval-oriented neural network (Neuro-Fuzzy BPN, called NF BPN) that is appropriate for assessing financial time-series data. Despite refining the flaws of point estimation in traditional neural network, it also reserved the non-linear predictive capability of neural network, to enhance the GARCH model in the insufficiency of independent variables. Also, it includes the price-quantity technical indexes and momentum of variances from financial time series as explanatory variables, to improve the details of evaluation in asset price/volatility of return rate. Besides, the trend of international financial liberation, loosening of regulation and open-up financial environment stimulate the emphasis of risk management. Value at Risk (VaR) model is able to quantify risk and wildly adopted by investors and financial institutes. Therefore, this research constructed the dynamic VaR model according to NF BPN, which may effectively capture the clustering phenomenon on asset prices/volatility of return rate, and further provide the investors and financial institutes as reference for risk management.This research takes the Taiwan Top 50 Tracker Fund (TTT 50) as empirical data to analyze, after comparing the empirical results and conventional GARCH model by evaluating their MSE, MAE and MAPE, the interval-base neural network model has the best prediction. In viewing the three features of accuracy, conservativeness and efficiency, the NF BPN has the better accuracy. The better conservativeness and efficiency of the GARCH model is meaningless under its weak accuracy. In general, this research suggests an innovative interval-oriented BPN model to evaluate the dynamic VaR of asset price/return and volatility, which may effectively apply the financial time-series data to assessing the possible loss under risk management. The physical momentum for price-quantity technical indicator and variance is also applied to enhance the explanatory capability of the model, to support the decision making process of the investors and financial institutes. Many other explanatory indexes are expected to include in the model to construct an excellent mechanism for risk management in the future, and the model may further be applied in other fields.