See summary poster here
Project Overiew
Accurate financial time series forecasts can assist investors in gaining a competitive edge over other participants in capital markets. Previous research yielded contradicting results pertaining to the superiority of machine learning (ML) over statistical models in time-series forecasting.
No empirical conclusion existed on what the most accurate model(s) were for forecasting stock market price movements over different forecast horizons. Existing studies either lacked comparisons to benchmarks, lacked multi-step-ahead forecasts, or did not use a large enough data set for statistically significant conclusions.
This study compared the forecasting accuracy of 20 different models on 403 time series of stocks/indices. These included four ML, seven statistical, and two benchmark models as well as seven different combinations of these models. The model forecast errors were assessed for 18 time steps into the future according to three different accuracy measures (sMAPE, MASE, and OWA).
No conclusion could be drawn on whether pure ML models always performed better or worse than pure statistical models. The top performing models included combinations of both pure ML and pure statistical models. The Naïve benchmark model outperformed all other models in this study for nearly all accuracy metrics and forecast horizons tested. This suggests that when forecasting monthly stock market price movements, no model from this study was more accurate than a simple random walk model.
Models Tested
-
Statistical Models
- Naive (random walk)
- Simple exponential smooting
- Holt’s linear trend
- Damped exponential smoothing
- Autoregressive (AR)
- Autoregressive moving average (ARMA)
- Autoregressive integrated moving average (ARIMA)
- Theta method
-
Machine Learning Models
- Long short term memory (LSTM) neural network
- Multi-layer perceptron (MLP)
- Support vector regressor (SVR)
- XGBoost
-
Various combinations of the above models
Acknowledgement
This project was funded by the DSI-NICIS National e-Science Postgraduate Teaching and Training Platform (NEPTTP). This project would not have been possible without the help from my supervisors Prof. Terence van Zyl and Dr. Farai Mlambo.