recap-regime-adaptive-portfolio

name	recap-regime-adaptive-portfolio
description	Regime-aware Continual Adaptive Portfolio management (ReCAP) — integrating continual learning into portfolio management via adaptive regime detection, policy libraries, and regime-gated policy combination. Accepted by KDD 2026. Activation: regime detection, portfolio management, continual learning, adaptive trading, ReCAP, market regime, policy library, regime shift.
category	finance

Context

Financial markets exhibit frequent regime shifts (bull/bear, high/low volatility) that render static portfolio optimization ineffective. ReCAP (Regime-aware Continual Adaptive Portfolio management) addresses this by combining regime detection with continual learning, enabling trading agents to accumulate and transfer knowledge across sequential market regimes. Paper: arXiv:2606.00143, accepted by KDD 2026.

Core Methodology

1. Adaptive Regime Detection: Segment historical market data into variable-length regimes using statistical change-point detection. Each regime represents a distinct market state (volatility level, trend direction, correlation structure).

2. Policy Library Construction: For each detected regime, learn and store a regime-specific policy vector. The policy library serves as a knowledge repository that grows as new regimes are encountered.

3. Regime-Gate Module: During live trading, a regime-gate adaptively combines policy vectors from the library based on the current market state. The gate weights determine which historical regimes are most relevant to current conditions.

4. Selective Continual Update: Only the regime-gate and the current regime's policy vector are continually updated. Historical policies are frozen to prevent catastrophic forgetting. This preserves accumulated knowledge while allowing rapid adaptation.

Implementation Steps

1. Collect historical market data (prices, volumes, features) and compute regime-sensitive indicators
2. Apply change-point detection algorithm (e.g., PELT, binary segmentation) to identify regime boundaries
3. For each regime, train a policy vector (e.g., portfolio weights) using the regime-specific data
4. Store policies in a dictionary: {regime_id: policy_vector}
5. In live trading:
   • Detect current regime using recent window of data
   • Query policy library for similar historical regimes
   • Regime-gate computes weighted combination of policy vectors
   • Execute combined portfolio allocation
   • Update only current regime's policy and gate weights

Key Results

• Outperforms rolling-window retraining and naive online fine-tuning across 5 real-world datasets
• Superior returns in long-term investment horizons
• Rapid adaptation to regime shifts without catastrophic forgetting
• Lower computational cost than full retraining approaches

Pitfalls

• Regime Detection Latency: Change-point detection operates on historical data; real-time regime identification has inherent lag. Use shorter detection windows for faster response but risk false positives.
• Policy Library Explosion: Too many detected regimes create an unmanageably large policy library. Consider merging similar regimes using clustering or similarity thresholds.
• Overfitting to Historical Regimes: Policies trained on specific regimes may not generalize to novel market conditions. Include regularization or policy smoothing.
• Gate Training Instability: The regime-gate module can become unstable during extreme market events. Use gradient clipping and learning rate scheduling.

Verification

• Compare ReCAP performance against: (1) rolling-window baseline, (2) naive online fine-tuning, (3) buy-and-hold benchmark
• Measure regime detection accuracy using held-out labeled regime data
• Verify that historical policy vectors remain stable (frozen) during continual updates
• Test adaptation speed after synthetic regime shifts in backtest

Activation

regime detection, portfolio management, continual learning, adaptive trading, ReCAP, market regime, policy library, regime shift, catastrophic forgetting, online learning, financial time series, dynamic allocation