Forecasting High-Frequency FX Return and Volatility : Evaluating Market Drivers and Probabilistic Models in the EUR/SEK Spot Market

Sammanfattning

Market makers in the foreign exchange market use high-frequency return and volatility forecasts to quote competitive prices and manage inventory risk. This thesis explores 30-second forecasting in the EUR/SEK spot market, with a focus on two questions: which market features drive return and volatility, and how much model complexity is needed to capture them. We compare six probabilistic forecasting models of increasing complexity, including discrete state models, linear and autoregressive variance models, and neural networks with temporal and distributional mixture extensions. Features include publicly observable order book data and SEB’s internal data. Models are evaluated using cross-validated log score, CRPS, calibration (PIT) analysis and directional accuracy, as well as a simplified trading backtest. The results show that order book imbalance and its change are the main external drivers of return, while market spread drives volatility. Models including internal features also consistently produce better trading performance and directional accuracy, most clearly during periods when those features are active, though these gains are within cross-validation variability. Model complexity improves performance more when addressing specific structural features of the data, than when simply increasing flexibility, although most differences are again within fold variability.