| Abstract [eng] |
The recovery of stock markets after a crash is a welcome development. The ability to estimate recovery length would help investors plan and achieve higher returns, protect against external risks and diversify better. Simultaneously, policy researchers and institutions could take more efficient fiscal and monetary policy actions during a crisis. However, predicting the length of the recovery process is rather challenging, as it can be impacted by various outside variables that change from crisis to crisis, such as the number of originators, the depth and duration of a crash, policy responses, an economic situation and psychological beliefs of individuals. This thesis used MSCI World Index monthly data between December 1969 and December 2022 and incorporated 19 popular features to forecast the index's recovery, defined as a return to pre-crash stock market peak price. For predictions, four neural networks-based models were chosen to be explored: multilayer perceptron, recurrent neural networks, long-short-term memory and gated recurrent unit. They were preferred over other machine learning models due to their ability to process sequential data. Statistical models were also not investigated due to prior research revealing their poorer performance attributable to stock market indices prices being chaotic, noisy and nonlinear. The final trained models, based on any historical 72 months information, are capable of providing price forecasts 72 months into the future. 72 months forecasting length was chosen due to research showing that crashes typically recover in less than six years. It was discovered that the multilayer perceptron model acquired the lowest MSE performance metric, equal to 0.09 on the entire test set, ranging from Great Recession to December 2022, with one hidden layer, eight nodes, a learning rate of 0.001, Adam optimiser and a ReLU activation function. The best MSE score measured only during the Great Recession period and equal to 0.03 was acquired by the multilayer perceptron with one hidden layer, 64 nodes, a learning rate of 0.001, Adam optimiser and a ReLU activation function. Regarding results from MAE and MAPE metrics perspective, multilayer perceptron models performed the best also for MAE and MAPE full test set data. The models did not beat the repeating time series baseline model only for MAPE and MAE values calculated during the Great Recession period. Yet, there was no drastic performance difference between the top baseline model and the runner-up neural network models in these cases. The performance of RNN, LSTM, and GRU models fell far short of MLP despite high hopes for them. Based on all MSE and MAE metrics and MAPE calculated on a full test dataset, the order of MLP being the best, followed by GRU, LSTM and then RNN was maintained. However, on MAPE calculated during the Great Recession period, GRUs scored a 13.35 % error, followed by RNNs at 15.06 %, MLPs at 15.99 % and LSTMs at 31.85 %. It was also noticed that the best models performed better during non-crash periods compared to crash periods. Interestingly, when forecasting recovery duration, the models usually commenced from a slightly larger or smaller price drop compared to the actual price but also typically displayed slower price-rising tendencies, eventually leading to an almost perfect match between the actual price and anticipated price. If one used the best MLP model during the lowest price point of the Great Recession, then the recovery point would have been predicted on point. GRU would have led to prematurely announcing recovery by a few months and using the best RNN and LSTM models, recovery would have been announced two years prematurely. Finally, the US consumer price index and the price of the MSCI World Index were the two factors that the results showed to be most crucial for making forecasts, followed by information on treasury bills and notes, the number of months the current crisis period has been ongoing, Williams % R, MACD and Stochastic % K technical analysis indicators. The least important indicators were the percentage price decline from the pre-crisis peak, the length of months that the present non-crisis phase has been continuing and the change in the gold price. |
