Forecasting with neural network during covid-19 crisis

• Thesis title in Czech: Předpovídání pomocí neuronových sítí počas krize covid-19
• Thesis title in English: Forecasting with neural network during covid-19 crisis
• Key words: Financial time series, ARIMA, GARCH, Neural Networks, forecasting
• English key words: Financial time series, ARIMA, GARCH, Neural Networks, Forecasting
• Academic year of topic announcement: 2019/2020
• Thesis type: diploma thesis
• Thesis language: angličtina
• Department: Institute of Economic Studies (23-IES)
• Supervisor: doc. PhDr. Jozef Baruník, Ph.D.
• Author: hidden - assigned by the advisor
• Date of registration: 05.06.2020
• Date of assignment: 05.06.2020
• Date and time of defence: 15.09.2021 09:00
• Venue of defence: Opletalova - Opletalova 26, O105, Opletalova - místn. č. 105
• Date of electronic submission: 27.07.2021
• Date of proceeded defence: 15.09.2021
• Opponents: PhDr. Jiří Kukačka, Ph.D.

References

1. Jasemi, M., Kimiagari, A.M. and Memariani, A., 2011. A modern neural network model to do stock market timing on the basis of the ancient investment technique of Japanese Candlestick. Expert Systems with Applications, 38(4), pp.3884-3890.
2. Chavarnakul, T. and Enke, D., 2008. Intelligent technical analysis based equivolume charting for stock trading using neural networks. Expert Systems with Applications, 34(2), pp.1004-1017.
3. Marshall, B.R., Young, M.R. and Rose, L.C., 2006. Candlestick technical trading strategies: Can they create value for investors?. Journal of Banking & Finance, 30(8), pp.2303-2323.
4. Chenoweth, T., Obradovic, Z. and Lee, S.S., 1996. Embedding technical analysis into neural network based trading systems. Applied Artificial Intelligence, 10(6), pp.523-542.
5. Hamid, S.A. and Iqbal Z., 2004. Using neural networks for forecasting volatility of S&P 500 index futures prices. Journal of Business Research, 57(10), pp.1116-1125
6. Roh, Tae Huyp 2007, Forecasting the volatility of stock price index. Expert Systems with Applications 33(4), pp. 916-22
7. Li, X., Deng. Z., Luo, J. 2009. Trading strategy design in financial investment through a turning points prediction scheme. Expert System with Applications, 36(4), pp. 7818-7826
8. Kristjanpoller, W., Fadic, A. and Minutolo, M.C., 2014. Volatility forecast using hybrid neural network models. Expert Systems with Applications, 41(5), pp.2437-2442.
Sources mentioned in the proposal but not constituting the core bibliography:
9. Brock, W. Lakonishok, J., and Lebaron, B., 1992. Simple technical trading rules and the stochastic properties of stock returns. The Journal of finance 47(5), pp. 1731-1764
10. Achelis, S.B., 2001. Technical Analysis from A to Z.
11. Burton, N., 2018. An Analysis of Burton G. Malkiel's A Random Walk Down Wall Street. CRC Press.
12. Horton, M.J., 2009. Stars, crows, and doji: The use of candlesticks in stock selection. The Quarterly Review of Economics and Finance, 49(2), pp.283-294.

Preliminary scope of work in English

Financial markets time series are highly nonlinear and stock prices can even seem to be completely random e.g. random walk. Traditional time series methods such as ARIMA and GARCH models are effective restricted on stationarity of a time series. Some time series even needs to be preprocessed by taking log returns. These issues therefore might decrease the effectiveness of the developed model in the live environment.

On the other hand, neural networks (NN) have become a major player in financial market predictions due to following properties (Jasemi, Kimiagari & Memariani, 2011):
1. Finance provides large datasets and neural networks utilize numbers in its nature
2. There is no need for input data distribution assumptions
3. The neural network can be further trained with new data available
4. Non-linear time series can be decoded
In the work done by Monica Lam (Lam, 2004), her NN utilized 16 financial statement variables and 11 macroeconomic variables for prediction. The NN that utilized solely one year’s or multiple’s years of financial data outperformed the minimum benchmark but did not outperformed the maximum benchmark. On the other hand, the NN utilizing the macroeconomic variables did not outperform minimum neither maximum benchmark values.

Technical analysis is a commonly used tool in stock trading. The work by Brock, Lakonishok and LeBaron validated the most popular and simplest trading rules, moving average and trading range break on Dow Jones index from 1876 to 1986. The returns obtained from these strategies were not consistent with four popular null models: the random walk, the AR(1), the GARCH-M and Exponential GARCH (Brock, Lakonishok & LeBaron, 1992). As technical analysis is gaining on traction in the recent years as there are questions regarding the rationality of investors (Marshal, Young & Rose, 2009). The technical analysis according shall reflect also the mood of the market regarding the human, political and economical events in its signals (Achelist, 1995). There are howevever also investors, that consider technical analysis as „voodoo finance“, thus disregarding its viablitiy such as Burton Malkiel.

In the work by Chavarnakul and Enke (Charvanakul & Enke, 2008), the technical analysis indicators, Volume Adjusted Moving Average (VAMA) and Ease of Movement (EMV), are combined with the NN and the newly combined model outperform the results of stock trading generated from the VAMA and EMV without the NN, from the simple Moving Average strategy (MA) and the Buy-and-Hold Trading strategy.

Japanese Candlesticks is an ancient trading strategy, which is based around opening, closing, high and low prices in intra-day trading. The combination of Japanese Candlesticks and NN yields astonishing results (Jasemi, Kimiagari & Memariani, 2011).

On the other view of the academic research, NN has been lately combined with AR, GARCH models in order to create a hybrid models to forecast the volatility of time series. As mentioned above, the financial time series are highly nonlinear econometrical models are often bounded by normality in some sort, which is limiting the performance of the forecast. By combining the NN with the traditional econometrical approaches, the goal is to improve the performance of volatility forecasts. In the work Volatility forecast using hybrid Neural Network models by Kristjanpoller, Fadic, Minutolo in 2014, the work tested hybrid NN-GARCH model to forecast the volatility of three Latin-American stock exchange indexes from Brazil, Chile and Mexico. The hybrid forecast significantly outperformed the pure methodology.

Another paper that utilizes the power of NN to combine with the econometrical approaches is the work done by Tae in 2007 in the paper Forecasting the volatility of stock price index. The work put into comparison the performance of pure NN model with hybrid models NN-ARCH, NN-GARCH, NN-EWMA, NN-EGARCH on the dataset of Korean Stock Exchange Market. By judging the performance on MAE, the models ranked as following: NN-EWMA > NN > NN-GARCH > NN-EGARCH.

Neural Networks were also used in forecasting the volatility, especially the deviation and direction in the work by Li, Zhidong, Jing in 2009. It uses the Turning Points Prediction framework. This paper proposed a trading strategy, which in the end resulted as a helpful tool for investors to make profitable decisions.

The motivation of my paper is to confirm the findings, that hybrid NN models in forecasting volatility outperforms vanilla econometric models in extreme situation as Covid-19. Secondly, in my paper, I will combine the NNs to create a trading strategy based on the forecast volatility provided by the NNs. This is to combine the findings of the two part of my literature review as listed before.

H1: The Neural Network forecast volatility model shall outperform pure GARCH volatility forecast during sudden shifts in the society such as the Covid-19.
H2: The trading strategies based on the NN forecast volatility shall outperform pure technical analysis in terms of return.

We will utlize common tools in financial market analysis, fundamental analysis and technical analysis combined with neural network to create the optimal portfolio and trading strategy.

Fundamental analysis is a method of evaluating the intrinsic value of an asset and analyzing the factor that could influence its price in the future. This is based mainly on financial statements and industry trends but they also account for external events and influences. The top-down approach evaluate the market first and then narrow it into sector, industry and finally a specific company. On the other hand, bottom-up approach starts with the specific stock and widens out to the all factors, that might affect the market thus consequently the price. Fundamental analysis is mostly used with shares, as one often use ratios such as price-to-equity, price-to-earnings, earnings-per-share multiples for fast comparison. Fundamental analysis bring a huge value in the terms of understanding the assets, which provides a long-term view of the assets on the market. The aim is to identify over- or under-valued assets, from which one can profit once a market correction occurs. However, fundamental analysis is very time demanding (apart from comparable ratio comparison), which might not be suited for fast decision making.

On the other hand, technical analysis is a method of examining and predicting price movement in the financial markets, using solely historical price charts and market statistic to determine the previous market patterns. From these identified market patterns, one is expected to be able to fairly predict a future price trajectory. Among favorite tools for technical analysis belong moving average, support and resistance level, Bollinger band and Japanese candlesticks. Technical analysis is popularly used among retail trades, as it fits the goal to gain in the short term, often in-day trades. As technical analysis is getting used more and more often, the trends and patterns are increasingly likely to be repeated thus helping the technical analysis to soar.

Neural Network is an element of Deep Learning, which is a subset of Machine Learning. Neural Network were developed as an inspiration of the human brain functionality. It constitutes of input layers, which are then processed by hidden layers with activation function and different weights and then yields the output layer. This is called a propagation process. A backpropagation process is a training process for the neuron network. Given that the output layer has a factual, real output, the output layer is then compared to the real output, adjustments are made to make the Neural Network more precise. As neural network is not dependent on linearity of provided data, it became a prominent tool in trading algorithm. Neural Network has many types such as Feedforward NN, Multilayer Perceptron, Recurrent Neural Network etc. One needs to then decide, which neural network seems to be the most fit for the given problem. For our task, we will be using Feedforward NN and Recurrent Neural Network.