Enhancing Forex Forecasting Accuracy with Hybrid Variable SetsEnhancing Forex Forecasting Accuracy with Hybrid Variable SetsEnhancing Forex Forecasting Accuracy with Hybrid Variable Sets

Data collection & preprocessing.
The authors gather two major data streams: (i) macroeconomic fundamentals from the U.S. and Euro-area (e.g., inflation, unemployment, government/external debt) and (ii) technical variables derived from daily EUR/USD price data (moving averages, oscillators, Ichimoku, Squeeze Momentum, support/resistance clusters, divergences, Fibonacci levels). Given the irregular cadence of macro releases, each sample day uses the latest published value and tracks "days since release" to encode recency.

Target definition and sequence design.
From the combined dataset of past features, they define a supervised classification task: will the next period's EUR/USD move up or down? They structure sliding-window sequences as input to an LSTM (or similar recurrent neural network) so the model can capture temporal dependencies.

Feature-set experiments.
Multiple variable-sets are defined technical-only, fundamentals-only, and hybrid sets combining both types (with varying inclusion of support/resistance and divergences). Each set is used to train an LSTM-based model with tuned hyperparameters (layers, window size, dropout, epochs). The goal: which input combination gives the best generalisation?

Evaluation and model selection.
Models are assessed via classification metrics (like AUC, accuracy, recall) and overfitting diagnostics (e.g., train-test AUC gap). The best models show acceptable test performance and low generalisation gap.

Trading simulation.
The selected model is deployed in a trading simulation on out-of-sample data. Signal thresholds based on predicted probabilities trigger long/short positions in EUR/USD, with realistic cost assumptions (e.g., spreads). The simulation reports win-rates, returns, and checks whether the system qualifies as a usable trading solution.

Core Findings

Construct a EUR/USD dataset (2012-2023) merging US/EU macro data (inflation, unemployment, debt) 
with daily technical indicators (SMA, EMA, Bollinger Bands, RSI, MACD, ADX, ATR). 
For each day, assign the latest published macro value and days_since_release. 
Then compute a binary label for the next 10-day movement direction. 

Train an LSTM with 4 layers, look_back=20 days, dropout=0.1. 
Output train/test AUC, AUC_diff, classification report, and model architecture summary.

Simulate EUR/USD trading on June 2023 March 2024. 
If model probability > 0.7 go long; if <0.3 go short. 
Use fixed 10-day hold, assume 1-pip spread. 
Compute total return, win-rates for longs and shorts.

Rich feature engineering pays: beyond standard indicators, structural features like clusters/divergences can improve timing.
Fundamentals still matter: even in high-efficiency FX markets, macro data contributes predictive power when incorporated properly.
Overfitting risk remains high: tracking generalisation (e.g., AUC_diff) is crucial.
Predictions alone are insufficient: embedding models into realistic trading simulations (costs, thresholds, out-of-sample) is what separates research from deployable systems.
Expect modest edges: in liquid FX markets, obtaining a stable edge is difficult; a modest but reliable gain is meaningful.
Shorting can show asymmetry: the study suggests the system performed especially well on downward moves, hinting at market structure nuances worth exploring in your own asset universe.
Not all popular features add value: e.g., Fibonacci levels may look good in textbooks but might not boost model performance when richer features are already included.