AI Investments created a complete trading solution, including a diversified way of signalling transactions, determining market conditions and managing exposition. The general idea was to apply DNN to reason about selected parameters, CNN to analyse visual patterns and LSTM to analyse the price & time sequence. For position size management and risk adjustment, the reinforcement algorithm with value network and MCST is used in a similar way as in Alpha Zero. Alpha Zero is the final version of the evolution of AlphaGo. The DNN and CNN are performing operations on the price movement data of different investment instruments in a specific time interval. Then, an additional dedicated DNN analyses the data generated by these networks to identify transaction signals.
Logical architecture of Investment Platform:
Neural networks for investment portfolio optimization
- DNN signal time & price – generates a transaction signal based on historical price data, seasonality, volatility and other price parameters
- CNN signal pattern-based image – generates a transaction signal based on the pattern recognition on the chart
- DNN range size – a network defining the range of traffic based on similar parameters as the first signalling network
- DNN market condition – a network defining general market conditions
- DNN control signal – a network defining a transactional signal based on data from the front networks.
Technical Implementation of Investment Platform:
Investment platform technical implementation
A unique element of the system architecture is the MCTS used to determine the best exposure of individual transactions providing its completeness. A tree that specifies different investment amounts in various transactions (generated on the basis of signals from the above DNN) is constructed and analyzed assuming the probabilities of positive results of individual transactions that are determined by previous neural networks. Based on that, the most optimal exposure strategy is evaluated.
The investment portfolio optimization platform includes:
- Neural networks for price, time, data and pattern recognition on the charts
- Reinforcement learning algorithm based on Alpha ZERO for position sizing, exposure and risk management
- Training pipeline – automated training and retraining of neural networks and reinforcement learning module
- Integration with the brokerage company for automated trading.
The described methods were chosen because they are most effective when it comes to generating trading strategies and improving investment returns It has been verified on real trading,
Walk-Forward Backtesting based on the results of Alpha ZERO and its
ability to learn from scratch. We are using the most advanced AI
approach available and applying it for investment portfolio optimization.
The portfolio optimization AI method works as follows:
- Financial time series pattern recognition networks trained for specific investment strategies
- Neural networks trained separately for buy and sell signals
- Boltzmann Machine and Restricted Boltzmann Machine networks
- Neural Turing machine
- Reinforcement learning algorithms.