beatnothing

beatnothing is a benchmarking framework designed to evaluate equity signals by answering a fundamental question: Can your model outperform a strategy that simply does nothing, considering costs and without foreknowledge of the future? This rigorous framework sets the bar for quant models by applying a simplistic trading approach—holding equal weights in a selected universe—while imposing identical transaction costs on both the benchmark and contestant strategies. The key output of this process is the Net Edge, which measures the difference in performance between a participant's strategy and the baseline strategy, highlighting the efficacy of different predictive models.

Overview

Each contestant in the beatnothing framework faces a challenging evaluation environment with defined rules:

1. One Engine: The model generates positions based on three pre-definition rules: long every name with a positive predicted value, long the top decile, or long the top decile while shorting the bottom, with specified exposure limits and fees.
2. Two Performance Benchmarks: Contestants are compared against a universal benchmark of equal-weighted positions and the RSP (the equal weight S&P 500 ETF) that operates without hindsight.
3. Transaction Costs: All implemented strategies incur predefined trading costs, ensuring a strong measure of real-world applicability.
4. Singular Scoring Metric: The evaluation is based on the Net Edge, calculated using Sharpe ratios, ensuring comprehensive analysis and comparison of results.
5. Frozen Submissions: Contestants' submissions are immutable, verified using SHA-256 hashing, thereby preserving the integrity of each model's performance record.

Season One Findings

The first season demonstrated that, despite the high performance metrics boasted by various models—including neural architectures and gradient boosted trees—none were able to exceed the benchmark's performance. All twenty-five submissions failed to clear the bar, revealing a critical gap between theoretical model performance and actual market outcomes. The results urged a reevaluation of certain strategies, such as:

• Cost Impacts: Significantly affect model performance rankings, highlighting that gross and net Sharpe ratios provide differing insights into a strategy's effectiveness.
• Model Generalization: Many predictors struggled when generalized across different trading contexts, emphasizing the importance of practical application over purely theoretical construct.

Features

• Comprehensive Data: The framework ships with historical price data from 1996 to 2026. This robust dataset allows contestants to validate their strategies against a realistic market backdrop.
• Benchmark Rigorousness: Evaluation is structured to discourage backtest overfitting through the implementation of statistical corrections and validations.
• User-Friendly Submission Process: New models can be introduced effortlessly, allowing users to test their strategies with minimal setup while maintaining a rigorous validation protocol.

Getting Started

With a straightforward installation process:

pip install beatnothing

Users can clone the repository and begin testing their equity signals against established benchmarks in under five minutes. The utility of this repository extends beyond a competitive platform; it offers detailed statistics and validation tools that can reveal critical insights about the effectiveness of different trading models in various market conditions. The project encapsulates a meaningful critique of current quantitative trading approaches, pushing for more stringent standards in the field.