In "vanilla" ML patters are usually static (oved decades cats remain the same). In contrast, the financial patterns are dynamic in their nature. They can evolve, disappear as well as reappear. Therefore, we have a problem of a moving target, which needs to be addressed properly if we want to produce strategies that can adapt to different states of the markets and, in this way, demonstrate a stable performance.Back to out AI Platform
METHODS WE APPLY TO SOLVE IT:
Time as Feature and Internal Data-Driven Adaptivity
Patterns existing on the market are dynamic. As soon as they are discovered, they become more and more utilized by the traders and, as a consequence, become weaker and even disappear. The dynamic nature of the patters raises the question of a proper training window. This question boils down to the question of an optimal balance between stochastic and systematic errors. If the training window is too small, we would have to few data points and, as a consequence, it an accurate description, or even reliable detection, of patterns will be impossible. On the other hand, if the training window is too large, we will see in the training patters that do not longer exist and, therefore, are not helpful or even harmful for the current performance of the strategy.
Within our approach we resolve the described problem by using time as an additional feature. In this case, the model itself can, in a data driven way, decide if a temporal split is necessary and, if it is the case, where exactly it should be done. This approach is more flexible than approach based on a training window of a constant size. Moreover, this approach is also more flexible because it does not force us to choose between two extremes: (1) use a fixed rigid patter for the whole available period of time, (2) only use the recent pattern and completely suppress the old ones. Instead of that the model can reuse an old pattern by correcting it or by combining the old patterns in a flexible way. This approach also provides an additional source of adaptiveness since a strategy developed in this way has an internal ability to perform a temporal extrapolation of existing trends.
Constant Monitoring of Strategies
Another way to address the problem of dynamic nature of patters is to perform a regular systematic model retraining with new available data. However, the problem of high noise-to-signal ration as well as peculiarity of financial KPIs requires us to be careful with a standard retraining and update of models. In more details, an inclusion of new data into the training will definitely lead to some, probably small, changes in model parameters. As a result, a regular model retraining would introduce some fluctuation of model parameters which, in its turn, would introduce some additional fluctuation of strategy’s returns and, in this way, worse Sharpe ratio.
To avoid the described problem, we adopt another approach. Namely, we update the model parameters only if we see, with a high statistical confidence, that the new version of the model really outperforms the old version of the model meaning that the observed difference in performance is not cause by noise and coincidence.