- The paper shows that integrating hybrid variable sets, combining fundamental and technical indicators, significantly enhances Forex forecasting accuracy.
- The study employs a rigorous LSTM-based framework with extensive hyperparameter tuning to detect profitable trading signals, especially for downward price movements.
- Empirical results from both fixed-horizon and dynamic backtesting confirm that hybrid models outperform technical-only approaches with minimal overfitting.
Integrating Hybrid Variable Sets in Cognitive Algorithmic Trading Systems for Forex Forecasting
Problem Statement and Motivation
The study investigates the predictive efficacy of integrating hybrid variable sets—comprising both fundamental macroeconomic indicators and technical analysis features—within cognitive algorithmic trading systems (ATS) for the EUR-USD Forex market. The research is motivated by the ongoing debate regarding the relative advantages of algorithmic trading versus human discretionary approaches, particularly the ability of cognitive ATS to assimilate and act upon the full informational context available to human traders. The central objective is to empirically determine whether such hybrid systems, operationalized through advanced LSTM-based architectures, can achieve statistically significant predictive superiority and trading profitability.
Feature Engineering and Variable Construction
The study employs a comprehensive feature engineering pipeline. Fundamental variables include sixteen macroeconomic indicators sourced from both the U.S. and Euro Area (e.g., HICP inflation rate, unemployment rates, net external debt, government debt metrics), aligned to a daily granularity by forward-filling their most recent values and including recency features. Technical variables span:
- Indicators and oscillators: SMA, EMA, Bollinger Bands, Ichimoku Cloud, RSI, MACD, ADX, Williams %R, ATR, KDJ Stochastic Oscillator, and Squeeze Momentum.
- Support and resistance levels: Derived by statistical clustering of local price extrema over rolling windows, yielding multi-level support/resistance variables per day.
- Fibonacci retracement and extension levels: Calculated dynamically using a 200-day window, with key levels appended as additional predictors.
- Divergence signals: Binary variables denoting convergence/divergence between price-action and Squeeze Momentum indicator trends, computed through slope analysis over recent local highs/lows.
The target variable is binary, indicating projected upward or downward price movement within a 10-day future window, using a composite statistic (D(n,h)) that includes maximal gain/loss and net change over the prediction horizon.
Modeling and Methodological Framework
The study formalizes the modeling challenge as a multi-objective, high-dimensional optimization problem, jointly tuning feature sets and LSTM hyperparameters. Ten model configurations were tested, ranging from price-only and technical-only to full hybrid models integrating all feature categories. Each configuration was subjected to an exhaustive randomized grid across 18 LSTM architectures, varying layer count (1/4/8), lookback window (20/30 days), and epochs (20/40/60).
Performance was evaluated using AUC (as the principal metric), accuracy, recall, detailed confusion matrices, and overfitting diagnostics (train/test AUC gap). The best models were further scrutinized through backtesting simulations, both with fixed and dynamic position management protocols.
Empirical Results
In-Sample Metrics
Aggregated results indicate that models incorporating fundamental data (particularly Models 2, 3, 5, 7, 9) display higher AUC values (0.62–0.65) relative to models based solely on technical variables (typically AUC ≤ 0.57). Optimal results were achieved with Model 2 (fundamental only), Model 3 (technical + fundamental), and Model 7 (technical, support/resistance, divergence, fundamental), with Model 7 displaying minimal overfitting (AUC_diff = 0.02) and high robustness across architectures.
Out-of-Sample Backtesting
Fixed-Horizon Protocol
In 10-day fixed-horizon simulation (post-June 2023), Model 7 produced the highest total returns and win rates (up to 100% win rate on short trades, 85.71% on long), substantially outperforming both technical-only and purely fundamental models. Notably, all models exhibited stronger predictive power for downward (short) signals than for upward (long) trades.
Dynamic Position Management
With dynamically managed positions, Model 7 again achieved perfect realized returns (all trades profitable) with limited trade frequency (4 trades for the out-of-sample period), demonstrating operational viability and resilience when facing real-market noise. Transaction cost analysis confirmed that results remained robust under competitive retail trading cost scenarios.
Feature Relevance
Inclusion of fundamental variables consistently improved predictive capacity, while Fibonacci levels did not yield substantial additive benefit, corroborating findings that some technical constructs provide minimal incremental signal under deep learning regimes with broad feature sets.
Theoretical and Practical Implications
The research provides strong evidence that cognitive ATS integrating hybrid variable sets, particularly macroeconomic fundamentals, can reliably surpass technical-only approaches in high-frequency Forex environments. Practically, this affirms the value of systematically encoding the full informational context accessible to human traders, thereby overcoming human cognitive bandwidth limitations.
Theoretically, the findings challenge the Efficient Market Hypothesis for currency markets at the operational timescales tested, supporting the thesis that statistical forecasting edges can be engineered through high-capacity, multi-modal neural architectures. The demonstrated capacity of LSTM-based cognitive ATS to profit in live simulations further underscores the transition of such approaches from academic concept to practical deployment.
Limitations and Future Directions
The investigation highlights several challenges: optimizing across vast hyperparameter and feature spaces, residual risks of overfitting, and the necessity for careful threshold calibration in position management logic. There is scope for exploration using alternative target definitions, integration of multi-timescale resolutions, inclusion of exogenous risk factors, and extension to a broader universe of currency pairs. Advanced hyperparameter optimization and real-time adaptive thresholds remain promising avenues. Systematic evaluation of feature selection strategies (e.g., information-theoretic or genetic algorithms) and benchmark replication across different markets and time periods are essential for further validation.
Conclusion
The study substantiates that carefully engineered cognitive ATS using hybrid variables, operationalized via rigorously optimized LSTM networks, can deliver measurable and exploitable forecasting advantages in Forex trading. The superior performance of hybrid models, especially those integrating fundamental macroeconomic information, sets a new reference point for predictive modeling paradigms in algorithmic currency trading. While implementation complexity is pronounced, the analytical capacity unlocked by these systems fundamentally exceeds practical human capability, marking a significant advancement in systematic trading research and application.