# Factor Mining - Discover and Mine Trading Factors **Last Updated**: 2026-03-17 **Version**: 1.0.0 ## How It Works ### Access the Factor Mining Lab Navigate to the Factor Mining page from the main menu. This lab provides AI-powered tools that automatically discover, generate, and evaluate alpha factors from historical market data using large language models. ### Select a Mining Mode Choose from three distinct mining approaches: Auto Discovery (AI generates novel factor hypotheses from scratch), From Report (extracts and formalizes factor ideas from uploaded PDF research reports), or Optimize Factor (takes an existing factor and iteratively improves its predictive power). ### Configure the LLM Provider and Model Select the AI language model provider and specific model to power the factor generation engine. Different models vary in creativity and analytical depth, which directly affects the quality and novelty of discovered factors. ### Set Mining Parameters Configure the loop count (number of discovery iterations), maximum duration, and hypothesis source. Higher loop counts allow the AI more attempts to discover high-quality factors but consume more computational credits and time. ### Upload Research Report (Optional) When using the From Report mode, upload a PDF research paper or factor analysis report. The AI extracts factor formulas, hypotheses, and logic from the document and converts them into testable factor implementations automatically. ### Launch the Mining Process Click Start to begin factor mining. The system enters a loop where the AI generates factor hypotheses, the backend implements them as executable code, and the evaluation engine tests each factor against historical data to compute IC, ICIR, and return metrics. ### Monitor Progress and Metrics Watch the real-time progress dashboard showing completed iterations, current IC and ICIR values, factor code previews, and evaluation results for each candidate. Factors that pass quality thresholds are automatically saved to your Factor Library. ### Review and Curate Discovered Factors After mining completes, review the full list of discovered factors ranked by IC and ICIR. Examine each factor's formula, backtest performance, and statistical significance. Promote the best candidates to your active Factor Library for use in strategy construction. > Factor mining uses AI-powered iterative discovery to generate alpha factor candidates. Each iteration consumes computational credits based on your subscription plan. ## Tips & Best Practices - Start with a small loop count (3-5 iterations) to evaluate the mining quality before committing to longer runs of 20+ iterations that consume significant credits - Factors with ICIR values above 0.5 show meaningful predictive consistency; prioritize these over factors with high IC but low ICIR, which may be unstable - Use the From Report mode with high-quality academic papers or institutional research to leverage proven factor hypotheses as starting points for AI refinement - After discovering promising factors, run them through Optimize Factor mode to iteratively improve their IC and ICIR scores through AI-guided parameter tuning ## Frequently Asked Questions ### What are IC and ICIR, and why do they matter? Information Coefficient (IC) measures the rank correlation between a factor's predicted values and actual future returns. ICIR (IC Information Ratio) is the mean IC divided by the standard deviation of IC, measuring prediction consistency. A factor with IC of 0.05 and ICIR of 0.5 is generally more reliable than one with IC of 0.10 but ICIR of 0.2, because consistency matters more than peak accuracy. ### What is the difference between Auto Discovery and Optimize Factor modes? Auto Discovery generates entirely new factor hypotheses from scratch using AI creativity, exploring a wide search space for novel alpha signals. Optimize Factor takes an existing factor as input and systematically varies its parameters, formula components, and lookback windows to maximize IC and ICIR. Use Auto Discovery for exploration and Optimize Factor for exploitation of known signals. ### How many computational credits does factor mining consume? Credit consumption depends on the number of iterations (loop count), the LLM model selected, and the complexity of factor evaluation. Each iteration involves an LLM call for hypothesis generation plus a backend computation for factor evaluation. Check your API Credits page for current usage and remaining balance. ### Can I stop a mining session and resume it later? Yes, you can pause a running mining session. Paused tasks appear in the Factor Task Manager under the Resumable tab, where you can restart them later. Factors already discovered before pausing are preserved and saved to your library. ### How do I know if a discovered factor is actually useful? Evaluate factors on three criteria: (1) Statistical significance - IC should be meaningfully different from zero with a t-statistic above 2.0, (2) Consistency - ICIR above 0.5 indicates reliable predictive power, (3) Economic intuition - the factor formula should have a logical explanation for why it predicts returns. Factors meeting all three criteria are strong candidates for strategy integration. ## Important Notes > AI-discovered factors must be validated on out-of-sample data before deployment. In-sample IC and ICIR metrics may overstate true predictive power due to data-mining bias. Past factor performance does not guarantee future alpha generation. --- Source: https://stratcraft.ai/help/factor-mining/