Loading...
Loading...
LeetQuant is an experimental framework for translating qualitative trading ideas into explicit, testable rules and observing how they behave across symbols, timeframes, and market conditions.
Much of this work is exploratory. The goal is not to prove that a given heuristic "works," but to understand where results are fragile, regime-dependent, or likely driven by noise rather than structure.
We utilize Large Language Models (LLMs) as translation layers to convert qualitative retail heuristics into standardized Python execution logic. Each strategy in our library is governed by strict, non-discretionary entry and exit rules, reducing discretionary interpretation so strategies can be tested consistently.
This translation step is imperfect and introduces its own modeling assumptions, which are a source of error rather than a guarantee of correctness.
The Leet Score is a composite ranking heuristic designed to surface strategies for further inspection. It is not an estimate of expected return, nor a claim of robustness. Different weightings or metrics would produce different rankings.
We do not rank by total return alone, as it fails to account for risk-adjusted performance.
Strategies are benchmarked against Buy & Hold (for the specific ticker) to ensure they are providing active value. Measures how much the strategy beats (or lags) a simple buy-and-hold approach over the same period. Capped at 30% outperformance = 100 points.
Higher Sharpe values indicate smoother return profiles under the sampled data. Capped at 5.0 for normalization. Unrealistic values (> 20) are filtered out.
Absolute return over the backtest period. While important, raw return is weighted less than risk-adjusted metrics to avoid rewarding strategies that take excessive risk. Capped at 50% = 100 points.
Percentage of profitable trades. Higher win rates contribute to trader confidence and smoother equity curves. Note: a low win rate can still be profitable with high risk/reward ratios.
We apply a logarithmic penalty to high-frequency signals to proxy for slippage and transaction costs. Fewer trades = better efficiency. More trades mean more execution costs, slippage, and time commitment. Formula: 500 / num_trades (capped 5-100).
Gross profit / gross loss. A profit factor > 1.0 means more money made than lost. Values above this range often warrant closer inspection. Capped at 5.0 for normalization.
We penalize strategies with significant peak-to-trough declines to reflect real-world capital preservation. A 20% max drawdown = -20 points. This discourages strategies that achieve high returns through excessive risk.
We use dynamic lookback windows tailored to the timeframe to capture recent "Regime Momentum":
Shorter lookbacks increase variance and the risk of overfitting. Longer lookbacks average across multiple regimes. The chosen windows reflect a tradeoff, not an optimal solution.
All backtests incorporate a standard 2-5 bps buffer to account for tiered commissions and slippage, ensuring that "paper profits" aren't lost to execution.
Our engine utilizes a strict (t-1) data availability model to prevent the inclusion of future information in signal generation.
We acknowledge the risks of Multiple Testing Bias. With 16,000+ daily backtests, "false discoveries" are inevitable. We experiment with out-of-sample checks and deflated Sharpe-style adjustments to monitor how quickly apparent performance decays.
Before calculating scores, we apply strict filters to ensure data quality:
Results shown are historical and highly sensitive to assumptions. Rankings change frequently and often do not persist. LeetQuant is a research tool, not investment advice.
Request research access to explore deeper diagnostics, trade history, and paper-trading behavior behind each strategy.