Effects of the validation set on stock returns forecasting

Abstract

Deep neural networks are potentially suitable tools for time series forecasting due to their ability to extract complex patterns of nonlinear data and their versatility in terms of models and applications. Even though they are powerful instruments and well-behaved approaches for certain tasks, they are sometimes surpassed by data complexity, and thus struggle to find an error that generalizes well enough on unseen data, especially in cases like times series forecasting for stock trading strategies. In this paper, the complex characteristics of time series are addressed by separating data by means of simpler yet more relevant distinctions in order to create a single model for every existing or created category, with a focus on model validation. This creates models which are trained on the same data, but validated for a particular class so that the models’ hyperparameters are specifically tuned to that class. Experiments on convolutional networks applied to the DJIA, Nasdaq and S&P 500 indices using volatility as a class or category indicator, have shown that it is possible to improve predictions after validating the model, obtaining the best model per the Model Confidence Set among different regression models on all time series datasets. Even the best and only model necessary for the DJIA and S&P 500 indices can be obtained at a significance value of 5% given that the level of volatility is known. The results highlight the importance of knowing the data and how to potentially separate them into simpler yet relevant classes. The results also reveal how model validation on different data is capable of creating models that better explain information just by tuning the model’s architectural hyperparameters, even though the models where trained on the very same data. This finding could be applied to any task requiring validation without modifying the training set, which is usually bigger and more expensive to obtain.