金融时序基础模型微调与回测方案生成器

You are a quantitative AI researcher specializing in financial time-series foundation models. I need a complete fine-tuning and backtesting plan for a **financial time-series foundation model**. ## Task Context - Asset class: [SPECIFY: stocks / crypto / forex / commodities] - Data frequency: [SPECIFY: tick / 1min / 5min / daily] - Prediction horizon: [SPECIFY: e.g., next 5 bars, 1 day ahead] - Base model: [SPECIFY: e.g., Kronos, TimesFM, Lag-Llama, or custom transformer] ## Please Generate: ### 1. Data Pipeline - Feature engineering: price, volume, volatility, order flow features - Normalization strategy (per-window z-score vs. global) - Train/val/test split with temporal awareness (no look-ahead bias) - Data augmentation techniques for financial data ### 2. Fine-Tuning Strategy - Which layers to freeze vs. train - Learning rate schedule (warmup + cosine decay) - Loss function selection (MSE vs. quantile loss vs. custom financial loss) - Regularization to prevent overfitting on regime-specific patterns ### 3. Backtesting Framework - Walk-forward validation setup - Position sizing rules derived from model confidence - Transaction cost modeling - Key metrics: Sharpe, Sortino, max drawdown, win rate, profit factor - Regime analysis: performance in trending vs. mean-reverting vs. volatile markets ### 4. Risk Controls - Maximum position limits - Drawdown-based circuit breakers - Model confidence thresholds for trade entry - Ensemble with rule-based fallbacks Provide concrete code snippets (Python) for each section where possible.