Financial Forecasting in the Data-Centric Era

Abstract

Deep learning (DL) has become an important tool in finance research, with applications ranging from asset pricing and portfolio allocation to risk management. However, empirical findings on its performance are often mixed, with DL frequently underperforming econometric or traditional machine learning methods. This work argues that a primary source of this inconsistency is the mismatch between the data-intensive nature of neural networks (NNs) and the data-scarce environments typical of many financial tasks. When trained at sufficient scale, NNs can deliver competitive or superior forecasts and can also inform theoretical research.

The first part of the thesis examines how data volume, model size, and architectural choice affect NN performance in financial forecasting. Using a dataset of more than 10,000 equities for volatility prediction, we find that performance gains from data scaling substantially exceed those from model scaling or architectural variation. We further demonstrate that a global training approach, which trains a single model across all assets, should be adopted as standard practice. The second part investigates the use of globally trained NNs as a tool for testing price discovery theories. We find that theory-driven variables effectively summarize the static state of the market but fail to capture temporal information in order flow. By training NNs on sequences of raw order book data, we provide empirical evidence against the Markovian assumption in traditional theory models. Finally, we show that NN performance can be further enhanced by hybridizing econometric and deep learning structures. We introduce the Realized Recurrent Conditional Heteroskedasticity (RealRECH) model, which augments the RealGARCH framework with a Long Short-Term Memory (LSTM) network.