| October 1, 2019
Project Overview
Literature about machine learning applied to automated trading most often reports results using a very crude metric: the sheer accuracy of the up/down prediction. Such a metric cannot guarantee actual trading performance: predictions have different financial consequences depending on the magnitude and direction of subsequent price movements. The goal of this project is to investigate whether it is possible to build a metric which would better reflect the results of backtesting.
The students first build simple deep-learning market prediction models, such as recurrent networks or one-dimensional convolutional networks, as a basis for evaluating candidate metrics. They then design a metric, or a group of metrics, that measures potential profitability during training for models that do not use reinforcement learning. Lusis provides a backtesting engine to check whether these measures reflect the performance of trading strategies. Finally, the metrics are used to select and improve the initial prediction models.
đź“Ž Presentation