One of the main problems in the field of financial investments is the ability to predict changes in asset prices, with emphasis on securities (securities / securities prediction) and then the creation of an "optimal" portfolio.
The DeepInvest project aims to develop innovative machine learning models and apply them in real-world, by creating real-world portfolios for securities returns and optimizing portfolio composition decisions. To this end, the proposal describes a program of research and parallel development of "practical" tools for forecasting performance and portfolio composition, which will lead to:
- Innovative machine learning models, well adapted to the above areas, which will allow the most efficient and faster (compared to the current literature) prediction of securities returns, based on recent developments in the field of neural networks, and early encouraging results by members of research teams, and
- automated portfolio management software, which will be aimed at investors and investment companies and which will achieve returns significantly higher than the market (eg stock indices), taking into account the profile of investors in relation to risk. Automated management will partially replace human management, be faster and less cost-effective, and scale economics.
The main goal of the DeepInvest project is to develop innovative methods for efficient stock price forecasting, based on new methodologies in the field of neural networks. For this reason, the following forecast optimization techniques will be explored and implemented:
- Neural networks based on LSTM (long short-term memory) units for predicting the returns of many securities per unit of time, without the need for training in each of them separately. This technique will allow the prediction of a large number of securities (instead of individual indices), taking as input of the network a given length of time series of shares, with the aim of multiple prediction of the returns of all securities.
- Investigation of "sequence-to-sequence" techniques, where two LSTM neural networks will be implemented, with one "handling" the encoding phase and the other the decoding phase. In the decoding phase, the array initializes the second network, to which additional data of a different nature than the first can be input. This effort will have the same goal, namely the high quality fast forecast for a large number of securities, using information related to the whole from historical data at different depths of time or with different frequency.
- Investigation of possible asymmetric loss function for the exploitation of "more" than the information of the predicted time series, incorporating both the sign and the size of the deviation of the predicted performance from the real one. With this technique, predictions that have an equal deviation from the actual performance but do not agree on a sign will be treated differently, so that the neural network "focuses" (and) on predicting the correct "movement" of the performance.
- Investigation of the recently proposed "attention" technique during training of neural networks, so that they take advantage of most of the input time series, with different focus weights on different samples of it. The successful implementation and improvement of the above technique will result in a significant reduction in training time and data needs.
- Implementation of optimization techniques using the expected portfolio returns, in order to determine the quantity ("weights") of the purchase of each security.