Paradox M5

Executive Summary

Strategy Overview

This is a mean-reversion trading system designed for XAUUSD operating on the 5-minute timeframe. The original research was conducted on the 1-minute timeframe in order to capture microstructure-level inefficiencies. However, significant differences in quote construction, tick aggregation, and feed quality across brokers produced inconsistent and non-reconcilable backtest results.   To mitigate these issues, the model was migrated to 5-minute bars, which significantly improves cross-broker data uniformity while preserving sensitivity to short-term price dislocations. The transition resulted in:

• Stronger parameter stability
• Improved cross-broker consistency
• A more reliable backtesting environment

Model Architecture

The strategy is constructed as an ensemble of decision units derived from quantitative market structure concepts.   Trade signals are generated through a coordinated decision process. Depending on prevailing market conditions, the system may require full consensus across decision units, or dynamically shift to a hierarchical decision structure. This adaptive mechanism allows the model to respond to changing microstructure dynamics while prioritizing long-term stability and controlled risk exposure.

Session-Aware Design

The model is session-aware. Each major trading session is treated as a distinct behavioral regime based on its historical characteristics.   Sessions are processed chronologically. When a new session begins, it supersedes the previous session’s governing rules. Only one position may be open at any time. If a trade remains active during a session transition, the risk management framework automatically switches to the parameter set associated with the new session. No overnight exposure is permitted. All positions are closed prior to market shutdown.

Research & Validation Methodology

To minimize overfitting to broker-specific microstructure artifacts, the strategy was validated across three broker models representing different execution environments:

• Broker A — ECN structure (tight spreads)
• Broker B — Market maker structure (moderate spreads)
• Broker C — Wide-spread broker model (large spreads and higher spread variability)

Each decision unit was independently optimized and evaluated across brokers. Rather than selecting peak-performance parameter combinations, the research focused on identifying stable clusters of parameters using performance clustering techniques. Only clusters demonstrating consistent behavior across brokers were retained. Any cluster that degraded significantly on even a single broker was rejected.

This process was repeated for each decision unit, each session and each broker until the full system architecture was assembled.   A structural out-of-sample boundary beginning 2024-01-01 was used to monitor performance degradation across brokers and sessions. Parameter clusters exhibiting meaningful deterioration during the OOS period were discarded.

Robustness Stress Tests

10,000 block-bootstrap simulations were performed to evaluate potential equity paths under randomized trade sequencing. All simulated equity paths generated a 0.0% probability of loss, indicating that system performance is not dependent on a specific historical trade order. To simulate missed executions and latency effects, trades were randomly removed from the sequence. Profit declines linearly as trades are removed, indicating that performance is evenly distributed across many trades rather than driven by isolated outliers. Additional transaction costs were applied to evaluate sensitivity to execution conditions against a baseline total profit of $29,934. Profitability remains highly stable, suggesting excellent resilience to moderate and severe spread variation or slippage.

Performance Summary

The system was evaluated from January 2020 through December 2025 (2190 days).

Account Performance

Annualized return (CAGR) during the period is approximately 26%.

Trading & Profitability Metrics

Risk Metrics

Out-of-Sample Behavior

A structural break was introduced at 2024-01-01.

The strategy remains highly stable under independent data.

OOS degradation analysis shows:

• Profit factor drop: −5.1%
• Win rate increase: +0.7%
• Average monthly profit increase: +43.8%

This behavior is consistent with highly robust systems surviving post-optimization reality.

Expectancy Analysis (Out-of-Sample)

To better understand the distribution of returns over time, the out-of-sample period was analyzed using both actual realized performance and bootstrapped Monte Carlo simulations. The objective of this analysis is to estimate the expected profit distribution across different time horizons while also quantifying potential upside and downside scenarios using Value-at-Risk style percentiles. Bootstrapped results were generated using 10,000 resampled sequences of the out-of-sample trade data.

Note: OOS Average Drawdown Duration is 13 days

The close alignment between actual mean returns and bootstrapped expectations across daily, weekly, and monthly horizons indicates that the observed out-of-sample performance is statistically consistent with the strategy’s underlying return distribution. Bootstrapped median returns closely match realized performance, indicating limited skew from extreme outliers. The 95% and 5% VaR bands provide realistic expectations for short-term profit and loss variability. Positive expectancy persists across all analyzed time horizons, supporting the stability of the strategy’s statistical edge.

Overall, the system’s profitability is not dependent on rare extreme events, but rather arises from consistent statistical expectancy distributed across many trades and time periods.

Risk Controls

Volatility Railguards

An optional module called Volatility Railguards filters trading activity based on a daily z-score of realized volatility. The internal volatility calculation closely mirrors the behavior of the CBOE Gold ETF Volatility Index. When enabled, trades are allowed only when volatility remains within ±2 standard deviations of its long-term mean. Extremely low volatility environments typically lack sufficient price movement. Extremely high volatility regimes introduce excessive execution costs and unstable price dynamics The module prioritizes capital preservation during abnormal volatility regimes, albeit at the cost of reduced trading activity. It can be particularly useful during major macroeconomic events such as:

• Nonfarm Payrolls releases
• Federal Open Market Committee decisions
• Jackson Hole Economic Symposium
• Interest rate announcements
• Major geopolitical developments

While this lowers drawdown risk, it also reduces return potential, and is generally intended for exceptional market conditions.

Dynamic Allocation

An optional position-sizing module called Dynamic Allocation adjusts the capital at risk per trade based on the strategy’s realized empirical performance rather than static risk inputs. The module calculates a rolling win rate over a user-defined lookback window (e.g., the last 30 trades). Risk is managed using a tiered step-function. When a session's win rate exceeds a defined upper threshold (e.g., 55%), risk allocation increases by a fixed step size for each additional percentage tier. When the win rate falls below a defined lower threshold (e.g., 45%), risk allocation is incrementally reduced to limit exposure. All adjustments remain within strict minimum and maximum risk parameters. This creates an algorithmic feedback loop that:

• Reduces drawdown velocity during unfavorable regimes
• Scales exposure only when a statistically verified edge exists

This calculation is strictly isolated per trading session (Asian, European, and American) to ensure that performance metrics in one market do not skew position sizing in another. Because the Dynamic Allocation module adjusts exposure based on session-specific performance, its effectiveness depends on the presence of persistent session-level regimes of outperformance or underperformance. However, statistical analysis of this strategy shows:

• Losses are not strongly clustered in time
• Underperformance does not persist within specific sessions
• The edge remains relatively stable across trading sessions

As a result, enabling Dynamic Allocation does not materially change the long-term performance profile of the strategy. In practice, the module produces only minor variations in equity growth and drawdown characteristics, with no statistically meaningful improvement or deterioration in profitability. For this reason, Dynamic Allocation should be viewed primarily as an optional risk management preference, rather than a core component of the system’s statistical edge.

Operational Requirements

• Spread widening, slippage, or latency may impact execution. ECN accounts are recommended to achieve best results.
• Session timing must align precisely with the broker’s server time, as the model relies on accurate session boundaries for regime classification.
• Take-profit and stop-loss levels should remain unchanged. These parameters were calibrated using the average session-specific daily range of gold over the past five years.