Strategy Arena methodology
The stock strategy pipeline starts with a shared data universe. Daily adjusted close series are pulled for each instrument, aligned to a common start date, and normalized to a base value for comparison across assets. The instruments cover index, commodity, and FX proxies that are widely traded and have long public histories. Only daily frequency is used in the core pipeline to reduce noise and to keep the same sampling for every strategy.
EA Arena methodology
The EA Arena uses an LLM-assisted pipeline to analyze MQL5 Expert Advisors. First, the MQ5 source code is parsed by GLM-5.1 into a structured JSON representation capturing indicators, entry/exit rules, parameters, and strategy type. Second, auto-convertible EAs are translated into Python backtrader strategies using a template-based code generator that maps MQL5 indicator calls to backtrader equivalents. Third, generated strategies are backtested on one year of H1 data for EUR/USD, GBP/USD, USD/JPY and gold with the same account for every EA: $10,000, 1:100 leverage, the EA's own lot size converted to contract units (100,000 per currency lot, 100 ounces per gold lot), one order at a time, and a broker-style stop-out when equity falls to 10% of the starting balance.
EAs that rely on MT5-specific features (custom indicators, DLL calls, or MQL5-only objects) are classified as manual-convert or mt5-only and are documented without backtesting. The convertibility classification ensures that only strategies with a reliable Python translation enter the backtest pipeline, keeping results comparable and interpretable.
Signal generation and execution
The daily strategy engine uses synthetic close-to-close accounting: the target decided from bars through close t earns the interval from close t to close t+1. This assumes allocation at the decision close; it is not a recorded next-open or next-day-close fill and may be optimistic about execution. The signal page separately shows the latest target and the previous target that earned the last observed interval. Converted EA simulations instead use next-bar-open execution on H1 bars. Neither timing convention alone proves that every underlying rule is free of look-ahead bias.
Transaction costs, slippage, and financing are not modeled in the baseline scores. The purpose is relative comparison across strategies, not a broker ready profit and loss statement. Readers should assume that high turnover models will perform worse in live conditions, and that liquidity constraints may invalidate some signals.
Evidence, observation dates and accounting corrections
The first evidence release covers six strategies on SPY, gold and bitcoin (18 pairs). The /evidence/ pages link actual configurations, code identity, daily derived records, simulated position changes and a standalone arithmetic verifier. Replay checks recorded arithmetic, not independent regeneration of decisions from a fresh provider download. Original inputs are retained privately; adjusted historical prices can be revised. Other strategies are not claimed to have complete evidence bundles.
Daily builds exclude bars dated today in UTC or later for every market. This deliberately delayed cutoff is not exchange-calendar aware; the actual observation date and build time are separate. From this release, WFO targets are stitched first and returns recomputed from the previous published target, including fold boundaries; skipped folds retain defaults. Earlier results could use a different candidate's preceding exposure. The evidence index records same-input accounting differences separately from changes to overlapping published history and retains correction snapshots.
Metrics and scoring
Stock strategies are scored with the Arena Score, a weighted average of percentile ranks (return quality, drawdown control, consistency, robustness and significance) within the same market and window. EA backtests report total trades, win rate, net P&L, maximum drawdown (both percentage and dollar), and an equity curve. The different scoring approaches reflect the different data frequencies and evaluation goals of each module.
Limitations and simplifications
The stock strategy pipeline assumes full allocation with no leverage and no short borrow constraints unless the strategy explicitly defines them. The EA backtest pipeline uses four instruments, one year of hourly data and a standard account, which may not match the instrument, timeframe, account size or broker settings the original EA was designed for. Both pipelines use public data sources that can differ from institutional feeds. Strategy results use walk-forward optimization only. Until 3 November 2026 the site kept the default parameters whenever they scored better over the full history, which applied to 106 of 576 strategy-and-market results; because that choice saw the whole sample, those results were slightly optimistic.
The MQ5→Python translation may simplify or approximate some indicator calculations. Custom indicators, multi-timeframe logic, and DLL-dependent features are intentionally excluded from auto-conversion. These constraints are part of the design and should be considered when interpreting backtest results.
Interpreting results
The methodology is intentionally transparent, but it is still a model. A high score or profitable backtest does not guarantee live performance, and a low score does not prove an idea is worthless. The output should be read as a research summary and combined with other due diligence such as liquidity analysis, execution modeling, and stress testing.
QuantArenas is updated with new strategies, new EA analyses, and new backtest results over time. Changes are logged in the changelog so readers can track when rankings shift and why a strategy may behave differently across releases.
All results are for research and education only. They are not financial advice and should not be treated as a recommendation to trade. Always validate a strategy in your own environment before taking any real market risk.
Last updated: 2026-11-06