Do artificial neural networks provide improved volatility forecasts: Evidence from Asian markets

The problem of forecasting stock market volatility remains a core issue in the empirical finance literature. While the impetus to earlier work began with the stock market crash of October 1987 (known as Black Monday), where twenty-three major world markets experienced substantial single day collapses,¹ repeated market events serve to highlight the importance of understanding volatility. Latterly, this includes the global financial crisis (GFC) that began in 2007, where the S&P 500 saw a weekly drop of more than 20%, and the COVID-19 pandemic in March 2020, which led to a dramatic fall in global equity markets. The DJIA index slumped more than 26% in four trading days, while the price of WTI crude oil fell into negative territory for the first time in recorded history. Global stock markets lost over US$16 trillion within 52 days. As history indicates, wide swings in stock markets lead to greater uncertainties that can be followed by the anticipation of a potential financial crisis. Thus, interest in modelling and forecasting financial markets has grown over the years with a desire to improve understanding of crises, tail events, and systematic risk.

The first general approach applied to this task within the academic literature is the genre of GARCH models (Engle, 1982; Bollerslev, 1986), while from a practitioner’s viewpoint the RiskMetrics variance model (also known as Exponential Smoothing) was introduced by JP Morgan in 1989. Subsequently, the volatility index (VIX), based on S&P 500 index options, was developed by the Chicago Board Options Exchange (CBOE) to measure stock market expectations in 1993. The VIX index is often referred to as a fear gauge by market participants, while similar indexes have been developed for a range of markets.

These models, and their extensions, receive notable attention from both financial academics and practitioners with a large amount of related published work. Nevertheless, combined with the characteristic constraints on historical volatility models and the growing technological transformation of financial markets, this suggests that new technologies might be better placed to provide improved volatility forecasts. Machine learning models (ML), based on Artificial Intelligence (AI) technology, have significantly improved in recent years and financial markets provide a fertile ground to examine the accuracy of AI-based volatility models against those traditionally considered.

There is no doubt that with the advent of the digital computer, stock market prediction has since moved into the technological realm. Moreover, the gains in computational speed have become increasingly important for banks, hedge funds and retail investors who are required to make investment decisions both within a short period of time and with large and constantly updating information sets. As news can now be processed quickly, there is a large increase in the volume of transactions, which generate further volatility. Thus, to overcome such restraints and effectively address today’s noisy, fast-paced and non-linear markets, AI techniques are proposed. Brav and Heaton (2002) argue that traditional market theories and methods may become incompatible with, and thus unable to model, the sophistication of modern financial analysis. Therefore, recent improvements in information technologies and the noted success of machine learning in pattern recognition given their flexibility and feasibility motivates researchers to use AI algorithms in stock price prediction (Bebarta et al., 2012; Heaton et al., 2017). One of the most successful examples of AI applications in financial markets is the performance of Medallion Fund with an average return of 66% over the last two decades (Kamalov, 2020). As these models are capable of learning non-linear patterns and functions, they are also demonstrated to be universal function approximators (Hornik et al., 1989; Kosko, 1994).

The present paper aims to contribute to the finance literature by improving the volatility forecasts of Asian stock markets using sophisticated neural network and machine learning techniques, against traditional benchmark econometric models, using both standard and economic-based evaluation measures. In doing so, the benchmark indices of ten emerging and developed Asian stock markets with the data running from 12 September 1994 to 05 March 2018 are utilised. Although a broad number of studies investigate stock market volatility using ML methods, there is a limited number that examine Asian markets, particularly emerging ones. Asian economies are blossoming in the last few decades by contributing almost 30% to the global economic output and making up over 40% of the world population (Jordan et al., 2017). In recent years, some researchers separately investigate volatility in major Asian markets, including Chen and Hu (2022) for China, Shaik and Sejpal (2020) for India, and Harahap et al. (2020) for Japan. However, the financial markets of emerging countries such as Indonesia, Thailand, Malaysia, and the Philippines, which together constitute 66% of the market capitalization of the ASEAN economies as of 2016 (Ganbold, 2021), tend to be ignored in volatility exercises. Furthermore, Asian economies, particularly emerging ones, differ from western and other developed economies in terms of cultural, financial, and institutional characteristics, which causes variation in forecast accuracy (Chen et al., 2010; Dovern et al., 2015; Jordan et al., 2017). In addition, the highly volatile behavior of these markets has the potential to impact regional and global stock markets through both the ‘leverage effect’ and idiosyncratic risk factors (Atanasov, 2018; Bouri et al., 2020). This further indicates the importance of generating more accurate and comprehensive forecasts for these markets.

Overall, this study aims to fill the practical gap on the optimal forecast model for Asian stock markets by evaluating several ANN (artificial neural network) models based on static, dynamic and supervised learning techniques and compare them with traditional methodologies, including GARCH and EGARCH models. To the best of our knowledge, only a limited number of studies compare standard ANN, Neuro-Fuzzy, and deep learning models within a wide range of emerging and developed markets. In contrast to previous work, this paper adopts and builds advanced neural network architectures for each selected model with improved learning rules and optimized hyperparameters. Moreover, not only do we conduct a comprehensive comparison between traditional forecasting methods and ANN models, but we also examine the economic implications of these models by assessing measures relevant for risk management practice. In doing so, this paper presents results of importance to both academics and market participants including investors and regulators.

The content of the paper is organized as follows. In Section 2 we briefly provide a review of the existing literature. In Section 3 we discuss the methodology, followed by Section 4 where we provide the data. In Section 5 we present the results and end the paper with concluding remarks in Section 6.

Volatility forecasting is a prevalent topic that has attracted scholars over the years. Engle (1982) addresses the volatility estimation problem in developing the ARCH model, which is considered one of the most significant developments in the empirical financial literature. It is followed by the GARCH model of Bollerslev (1986) and subsequently, a number of GARCH extensions. The volatility literature also saw the development of alternative approaches, including, linear regression models, hybrid models, the support vector regression, fuzzy logic, genetic algorithms and artificial neural networks.

One of the earliest studies on machine learning forecasting is Yoon and Swales (1991) who examine the stock market data of 58 widely followed companies in the Fortune 500. Their findings reveal that neural networks significantly improve stock price predictability compared to conventional methods. Donaldson and Kamstra (1996) improve the applicability of the ANN approach using time series data on four developed stock markets. They conduct out-of-sample forecasts and reveal that ANN is superior in terms of forecasting stock market volatility compared to traditional linear models given its flexibility with complex nonlinear dynamics. Furthermore, Ormoneit and Neuneier (1996) apply ANN models on the German DAX index by using minute data for the month of November 1994. They compare the Multilayer Perceptron method (MLP) with the Conditional Density Estimating Neural Network (CDENN) and report that improved predictions can be achieved by more complex architectures that target noise in the stock market. In light of this study, Kim and Lee (2004) propose the feature transformation method based on the Genetic Algorithm (GA) model and compare it with two conventional neural networks. The results indicate that the GA method improves prediction capability for financial market forecasting compared to the conventional ANNs. Altay and Satman (2005) implement ANN methods on the Istanbul Stock Exchange by using daily, weekly, and monthly data. They compare out-of-sample forecasting results with linear regression models and report that ANN is superior only for weekly forecast results, while underperforming for daily and monthly data. Cao et al. (2005) study ANN methods to predict firm-level stock prices that trade on the Shanghai stock exchange. They compare univariate and multivariate ANN models with linear models, with the results indicating superiority of the neural network models in predicting future price changes. In contrast, Mantri et al. (2014) investigate the two Indian benchmark indices (BSE SENSEX and NIFTY) from 1995 to 2008 by comparing GARCH, EGARCH, GJR-GARCH, IGARCH and ANN models. The authors report that the prediction ability of the ANN model offers no improvement over statistical forecast models.

A number of studies investigate the performance of a different class of ANN and hybrid models. Roh (2007) proposes a hybrid model between ANN and time series econometric models for the KOSPI Index, with forecast results supporting the accuracy of the hybrid model for volatility. Unlike Roh (2007), Guresen et al. (2011) analyse daily NASDAQ returns but find that hybrid models are not as successful as standard ANN models. Kristjanpoller et al. (2014) propose ANN-GARCH hybrid models to predict three emerging Latin American stock markets and conclude that hybrid models improve prediction ability over conventional models. Further studies relating to hybrid models are undertaken by Rather et al. (2015), Mingyue et al. (2016), Kim and Won (2018), and Hao and Gao (2020).

Adebiyi et al. (2012) combine technical and fundamental analysis with ANN and provide results suggesting that the combination improves prediction, consistent with the findings of Yao et al. (1999) and later supported by Sezer et al. (2017). However, Namdari and Li (2018) report mixed results in terms of forecasting ability of an integrated ANN with fundamental and technical analysis models. Of note, the results show that the integrated model works well when the stocks have an upward trend.

Several researchers experiment with neuro-fuzzy and neuro-evolutionary methods in stock market forecasting exercises, which are believed to have a combination of advantages through ANN and fuzzy logic. Quah (2007) uses DJIA index data spanning from 1994 to 2005 to compare the applicability of MLP, ANFIS and GGAP-RBF models. Using several benchmark metrics, including generalize rate, recall rate, confusion metrics and appreciation, the study shows that ANFIS provides more accurate results while GGAP-RBF underperforms in all selected criteria. In a similar vein, Yang et al. (2012) find that a fuzzy reasoning system can be used to predict stock market trends. Li and Xiong (2005) argue that neural networks have limitations in dealing with qualitative information and suffer from ‘black box’ syndrome, proposing a neuro fuzzy inference system to overcome these drawbacks. The Shanghai stock market is chosen for prediction where they find the suggested fuzzy NN is superior to standard NN methods. Mandziuk and Jaruszewicz (2007) propose a novel neuro-evolutionary method to predict the change in the daily closing price on the DAX index. The results reveal that the proposed model produces high accuracy for the market in both directions. García et al. (2018) implement a hybrid neuro fuzzy model to predict one-day ahead direction of the DAX Index. They conclude that the integration of traditional indicators with ANN may enhance predictive accuracy of the model, although it may also generate noise in the prediction model. Further discussion is reported in D’Urso et al. (2013), Vlasenko et al. (2018) and Chandar (2019).

More recently, Luo et al. (2018) discuss that the predictive capabilities of deep learning models are lesser compared to other ANN algorithms. Although D’Amato et al. (2022) demonstrate the suitability and capability of deep learning models on the inherently complex and chaotic crypto market. Koo and Kim (2022) propose a new model by combining GARCH and LSTM models with a volume-upped (VU) distribution strategy. They conclude that the proposed model improves forecasting performance by 21.03% compared to standalone deep learning models. Similarly, Ahamed and Ravi (2021) investigate the shortcomings of a deep learning network by focusing on the optimization problem and address the issue by using swarm intelligence algorithms. Whereas Yang et al. (2020) reveal that the predictive power of genetic algorithms is better than the swarm optimization.

The above discussion demonstrates that the present state of the literature does not suggest a clear superiority either within the different ANN models, or over conventional forecasting methods. As Ravichandra and Thingom (2016) and Chopra and Sharma (2021) discuss, AI models do possess superior capabilities and the potential for more accurate volatility forecasts and thus, worthy of further research. This paper builds upon the research in the current literature comparing the volatility forecasting capabilities of ML models to traditional models and extends the literature by evaluating a wider set of ANNs and utilising risk management measures and so developing the economic implications.

In this research, ten Asian countries and their widely accepted indices are chosen. These markets are: the Nikkei 225 Index (NIKKEI) from Japan, the Hang Seng Index (HSI) from Hong Kong, the Korea Composite Stock Market Index (KOSPI) from South Korea, the Taiwan Capitalization Weighted Stock Index (TAIEX) from Taiwan, the Straits Times Index (STI) from Singapore, the SSE Composite Index (SSE) from China, the PSE Composite Index (PSE) from the Philippines, the Stock Exchange of Thailand Index (SET) from Thailand, the Kuala Lumpur Composite Index (KLCI) from Malaysia, and the Jakarta Stock Exchange Composite Index (JCI) from Indonesia.

The sample period spans from 12/09/1994 to 05/03/2018, with Table 1 reporting the selected markets (and indices) and sample sizes (including out-of-sample forecast period) for each market, respectively. This period is selected based on data availability and covers important financial events such as the Asian financial crisis of 1997-98 and the global financial crisis of 2008. Table 2 presents the key descriptive statistics for the total data sample for each index. The mean fluctuates between 0.0047 and 0.0448 for daily returns. Indonesia outperforms other markets while the Thai stock market performs worst. The return distribution is not symmetrical, with the series exhibiting skewness. The values in Table 2 suggest that half the selected markets exhibit negative skewness, with the other half indicating positive skewness.² The results also suggest the presence of excess kurtosis, which indicate a larger number of extreme shocks (of either sign) than under a normal distribution. Of further note, China has the highest maximum value, while Singapore and Taiwan have the lowest maximum values. The greatest single-day increase is in China’s SSE of 26.99% and the largest decline occurs in Malaysia’s KLCI with -24.15%. Singapore’s STI and Taiwan’s TAIEX Indices have the smallest gap between daily minimum and maximum values of -8.70% and 7.53% and -6.98% and 6.52% respectively. This result indicates lower volatility compared to others, which is also seen in the standard deviation values.³

Table 3 presents the forecasting performance results for daily return series based on the Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE), Quasi-Likelihood (QLIKE) and Mean Squared Error (MSE) measures. The out-of-sample forecasts are obtained using the ten ANN models and four benchmark models. The overall results suggest that the benchmark models provide superior forecasts based on the MAE criterion for seven of the ten indices, with the only exceptions of STI, KLCI and JCI indices. The result for the KLCI index is consistent with the study of Yao et al. (1999). According to the MAPE criterion, ANN models clearly outperform the benchmark models. Notably, the RNN, RBFN and PNN models provide the lowest MAPE values across multiple indices. In terms of the RMSE loss function, the EGARCH model achieves the best results in KLCI and TAIEX indices, whereas the GARCH model performs the worst among all. LSTM model tends to provide more accurate forecast results compared to other models. This contrasts with the work of Selvin et al. (2017), although supports the findings of Chen et al. (2015) and Nelson et al. (2017). The QLIKE and MSE error criteria find substantial support for the prediction power of ANN-based models with the only exception of STI, KLCI and TAIEX indices, for which they provide either mixed results or favour traditional forecasting models. The adaptive and coactive network-based hybrid models of ANFIS and CANFIS indicate the lowest prediction errors specifically in HANG SENG, TAIEX and PSE indices, which supports Chang et al. (2008), Boyacioglu and Avci (2010), and Kristjanpoller and Michell (2018). The comparative predictive performance of standard NN, neuro-fuzzy and deep learning models indicate robust results compared to conventional methods for more occasions than the reverse. More specifically, the LSTM provides superior forecasts for six of the ten markets based on the MSE criterion, which justifies its preferential role in long-term time series predictions given its memory cell properties (Kim and Kang, 2019). Other deep learning models, such as RNN, MLP and RBFN, are superior in three, three and four occasions respectively. In addition to the findings of Yap et al. (2021) on using deep learning models for predicting short-term movements and market trends in Asian tiger countries, the present results show that deep learning models are preferred in forecasting a wider range of markets. Furthermore, neuro-fuzzy models are favoured specifically for the NIKKEI, HANG SENG, SSE, TAIEX and PSE indices, despite it clearly underperforming for the remaining markets. Although Atsalakis et al. (2016) state that Neuro-fuzzy models are more preferred for turbulent times and shorter-term predictions given their rapid learning capabilities, these results show that neuro-fuzzy models also offer promising results over longer-term periods. GFN, MFN and PNN models indicate outperformance in seven, five and two occasions respectively. Notably, the MFN is clearly preferred for KLCI index where four out of five losses indicate preference. The GFN model reports its lowest errors based on RMSE, QLIKE and MSE for JCI index. The PNN model is the weakest among all ANN models where it is only preferable based on MAPE criterion for TAIEX and HANG SENG indices. This result supports the view of Chen et al. (2003) for TAIEX index where PNN also produces enhanced predictive power compared to parametric benchmark. However, as indicated by Wang and Wu (2017), the overall weaker performance of PNN might be due to its high computational complexity in the standard architecture which causes difficulties in the estimation of parameters.

To provide some further understanding of the nature of the results, we consider the cumulative MSE and QLIKE plots for the ANN and GARCH models. The cumulative plots allow us to consider whether any forecast improvement occurs consistently over the sample period or whether it is associated with a particular date or event. To summarise the information across the nine different ANN models, we use the combined forecast series as defined by equation (23), while the same approach is undertaken to obtain a combined GARCH and EGARCH forecast.

Figure 1 presents the comparison of the cumulative MSE and QLIKE error functions over the out-of-sample period for the combined ANN and GARCH models for each index. One clear characteristic across the graphs is the jump associated with the 2007-2008 crisis in almost all markets and which is reflected in the MSE loss function more noticeably than the QLIKE error criterion. This also clearly highlights that large volatility increases that occur during turbulent times present difficulties in forecasting and applies to both ANN models and GARCH models. To consider a further example of the same effect, we can observe a jump in the Shanghai Composite Index during the 2015-2016 Chinese Stock Market turbulence.

In considering the relative performance of ANN versus GARCH models, we can see that for the cumulative MSE graphs, the two forecast series track each other closely with GARCH approach typically offering a smaller value. One notable exception is for the STI where the GARCH model is clearly preferred throughout. Moreover, the forecast improvement with the GARCH model largely arises after the financial crisis period, with this being most obvious for the KLCI. Considering the cumulative forecasts for the QLIKE measure, we can see that the ANN approach consistently outperforms the GARCH model, with the except of the STI. Moreover, this forecast improvement is not associated with a particular event, as we see with the MSE but over the whole forecast sample the ANN model shows an improving outperformance i.e., the gap between the two cumulative series expands. Furthermore, for most of the series, we see the ANN QLIKE function exhibiting greater stability compared to the GARCH model where there is a continuous increase in QLIKE values.

Following the studies of Hansen et al. (2011), Wang et al. (2016), and Liu et al. (2017), we also consider Model Confidence Set (MCS) test with a confidence level of 75% which allows us to compare the given model set in MCS framework with a p-value larger than 0.25. Table 4 exhibits the MCS test for both MSE and QLIKE metrics based on the out-of-sample forecasting results. The bold values in the table denotes the optimal model chosen by MCS, while the test also considers the number of other models with EPA at the given confidence level. The corresponding results of the MCS test indicate the ANN class of models are significantly better than the benchmark models. Specifically, LSTM model is preferred on five occasions based on the MSE criterion and three occasions based on the QLIKE loss function. Moreover, the QLIKE loss function supports the superiority of MFN model in five markets, while traditional methods are eliminated in most cases. In summary, the MCS test results confirm the superiority of ANN models over the benchmark models of GARCH, EGARCH, MACD and Naïve models.

Table 5 presents the daily VaR and Expected Shortfall statistics as well as the corresponding test results. Examining Table 5, the lowest average VaR failure rate at the 1% level is mainly achieved by the hybrid models of ANFIS and CANFIS, while the benchmark models of GARCH and EGARCH report lowest values in KLCI and SET indices. The PNN model provides the preferred average failure rate for KOSPI index, while the RBFN and PNN models are preferred for the SSE index. In contrast, the LSTM, RNN and MLP models fail to provide minimum VaR rates for any of the selected indices and for which they tend to underestimate potential risks. As recently proposed by Basel Committee in 2017, there is a move regarding quantitative risk measures from VaR to ES (Expected Shortfall). In forecasting ES, the MLP model is preferred at 1% and 5% levels for the SSE, PSE, STI and HANG SENG indices. Furthermore, the RBFN, MFN and PNN models are preferred in both confidence levels for NIKKEI, KLCI and KOSPI indices. Accordingly, it can be inferred that the ANN models are the most suitable across all competing models in terms of Expected Shortfall at all selected confidence levels. The accuracy and reliability of the VaR forecasts are also tested as proposed by Basel Ι and Basel ΙΙ. Based on the tests of Kupiec, Christoffersen and DQ, the results report that none of the models reject the null hypothesis of expected VaR violation (Kupiec’s unconditional coverage test), the independence exceptions of VaR (Christoffersen’s conditional coverage test), and violations of VaR occurred correlated (Dynamic Quantile).

Overall, the results highlight the accuracy of the ANN class of models for volatility forecasting both in terms of statistical measures and economic, VaR and ES, metrics across a range of Asian stock markets. Notably, while there are exceptions, the results, similar to Zhang et al. (1998) and Cao and Wang (2020), suggests that the class of ANN models outperforms traditional forecasting methods across statistical and economic measures.

Volatility forecasting is essential for both practitioners and policymakers to enable them improve decision making and portfolio building, especially during periods of financial turbulence. This paper evaluates different Machine Learning methods in forecasting the volatility of ten Asian stock market indices, with the results compared against benchmark models. The empirical results for ANN models are promising. Out-of-sample forecast evaluation shows that ANN models are preferred for each index compared to the GARCH and EGARCH models. Notably, the results show that neural network prediction models exhibit improved forecasting accuracy across both statistical and economic-based metrics and offer new insights for market participants, academics, and policymakers. Although, it should be noted that the GARCH models do perform well across some individual series.

The contribution of this paper to the field of empirical finance is three-fold. First and foremost, this study explores key relevant machine learning models to address the problem of financial volatility forecasting. Previous studies tend to evaluate small sets of neural network methods. Using a wider range of ANN architectures has various advantages. For example, in stock market prediction exercises, recurrent ANNs are recommended due to their memory component features that increase prediction accuracy. Second, comprehensive performance measures for model evaluation are utilized, namely, both a range of statistical measures (RMSE, MAE, MAPE, MSE, QLIKE, and MCS) and economic-based ones (VaR and ES). Third, a wide range of Asian markets were studied in order to have an in-depth examination of an extended set of volatility models across markets that are less studied.

To extend the study, additional research could explore a further diverse set of ANN architectures. For example, according to Partaourides and Chatzis (2017), further regularization methods may increase the capacity of the machine learning systems. Moreover, hidden layers can be extended beyond two, more data frequencies can be added, and alternative input variables and activation functions can be studied. The value of such novel developments remains to be examined in future research endeavors.