Comparing Direct, Recursive, and Multi-Output Forecasting for EURUSD
Summary
The article introduces multi-horizon forecasting and compares three approaches: training a separate model for each forecast step, iteratively feeding predictions from a single model back into itself, and predicting several horizons together with a multi-output model. It describes preparing EURUSD hourly OHLC data and binary candle-direction labels, then training LightGBM classifiers for separate future steps. The forecasts are exported in ONNX format for use in an MQL5 Expert Advisor, which displays predicted signals for upcoming bars. The author also outlines recursive linear regression and a neural network with multiple outputs.
For the direct LightGBM example, reported test accuracy varies by horizon, with the second step performing best at about 55%; the other listed results are near or below chance. These are limited classification results on a small sample of hourly data, not evidence of a profitable trading strategy. The article explains tradeoffs: direct models avoid cascading forecast errors but can disagree across horizons and require more models, while recursive forecasts can accumulate error. No deployment profitability or robust out-of-sample evaluation is established.
Key ideas
- Direct forecasting trains an independent model for each future horizon, allowing horizon-specific fitting at added maintenance cost.
- Recursive forecasting reuses one-step predictions as inputs for later steps, so forecast errors can compound.
- Multi-output models predict several future values together and can represent relationships among horizons.
- The EURUSD example trains LightGBM classifiers on hourly candle signals and deploys their ONNX models in MQL5.
- Reported accuracy varies across horizons and does not demonstrate trading profitability.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.