q.label Roadmap

Open work for future-aware target construction. The candidates below draw from methods Marcos López de Prado developed, co-developed, or popularized in his financial-machine-learning research.

Event discovery

• Structural-break event filters: Chu-Stinchcombe-White CUSUM, Chow-type Dickey-Fuller, and SADF/explosiveness tests
• Volatility-normalized CUSUM thresholds with explicit lagging rules so an event threshold uses only information available at that timestamp
• Information-driven event times from tick, volume, and dollar imbalance bars and run bars; bar construction belongs in q.data, while q.label consumes the resulting event index
• Event-filter diagnostics comparing clock-time sampling with information-driven sampling, including event density and regime stability

Label construction

• Optional high/low intrabar barrier touches with an explicit same-bar tie policy
• Event-specific asymmetric barriers and dynamic vertical horizons derived from point-in-time volatility or market regimes
• Optimal profit-taking and stop-loss calibration for mean-reverting and momentum processes without brute-force backtest parameter searches
• Trend-scanning extensions for irregular timestamps, session-aware horizons, robust regression, and multiscale label consensus
• Rare-label pruning with transparent minimum-frequency rules and a report of discarded observations
• A first-class meta-labeling workflow that joins primary-model side, predicted probability, barrier outcome, and secondary binary target without losing event-lifetime metadata
• Label-stability surfaces across barrier multipliers, horizons, and target widths to reveal targets that depend on a narrow parameter choice

Uniqueness and sample weights

• Sparse indicator-matrix construction showing which observations belong to each event lifetime
• Sequential bootstrap sampling based on average uniqueness, with a seeded and reproducible draw API
• Linear time-decay weights that preserve a configurable fraction of older observations instead of silently deleting them
• Piecewise and nonlinear time-decay schedules
• Composable weights combining average uniqueness, concurrency-attributed absolute returns, time decay, and optional class balancing
• Uniqueness-aware train/bag sampling adapters for q.model, while keeping estimator implementations outside q.label

Diagnostics and scale

• Label-distribution, barrier-touch, event-lifetime, concurrency, and uniqueness diagnostics under q.plot
• Purging and embargo metadata export for q.dataset, including the maximum label end time required to prevent overlapping train/test paths
• Synthetic Ornstein-Uhlenbeck, momentum, and regime-switching paths for measuring label recovery, stability, and class balance against known dynamics
• Scalable chunked execution and sparse interval operations for very large event sets