跳转到内容
Crypto Quant Docs

退役策略记录

Retired Strategies — Reasons & Lessons

Section titled “Retired Strategies — Reasons & Lessons”

2026-04-22 update: 代码已从 strategies/legacy/ + configs_legacy/ 删除。 本文档是仅存的退役记录 —— 每个条目说明 为什么不用 + 教训是什么。 想复活请严格按末尾 “Restoration Protocol”。

The active deployed stack lives in strategies/. See HONEST_TREND_REPORT.md for the full validation report.

TL;DR — Patterns That Fooled Us

Section titled “TL;DR — Patterns That Fooled Us”

Every strategy here passed at least one of these “looks great” tests and still failed in reality:

  1. In-sample PF > 3.0 → almost always overfit, 90% of the gap between IS and OOS is noise.
  2. Calmar > 20 → red flag, not a feature. Means the sample is too small to have hit a real drawdown yet.
  3. p-value < 0.0001 in-sample → statistical significance on IS data is circular reasoning.
  4. Single OOS window profitable → one lucky quarter can hide 4 losing quarters. Always walk-forward.
  5. “Contrarian at extremes” LLM prompts → Claude sets contrarian=true 96% of the time regardless of prompt. Signal is random noise with extra steps.

Sentiment / ML Family (overfit cluster)

Section titled “Sentiment / ML Family (overfit cluster)”

SentimentTrend.py — live-mode sentiment strategy with LLM signals

Section titled “SentimentTrend.py — live-mode sentiment strategy with LLM signals”
  • Why retired: PF 6.86 in-sample collapsed to ~1.05 OOS.
  • What looked good: combined hand-crafted FnG gate + LLM sentiment + reflective memory + BTC cycle indicators → 200+ features.
  • What actually happened: feature importance analysis showed LLM signal contributed 0%. Redundant with FnG. Most “features” were noise that happened to correlate with price in 2024 bull market.
  • Lesson: adding features ≠ adding edge. Always ablate.

SentimentTrendBT.py / SentimentTrendBT.json

Section titled “SentimentTrendBT.py / SentimentTrendBT.json”
  • Why retired: Backtestable version of SentimentTrend.py. Used historical FnG + reconstructed LLM outputs. Tested 6 variants (v1–v6), none had OOS edge.
  • Lesson: historical sentiment reconstruction is a leaky abstraction.

SentimentUltimate.py

Section titled “SentimentUltimate.py”
  • Why retired: Hand-crafted + FreqAI LightGBM classifier hybrid. In-sample PF 6.86 → OOS PF 1.05. Textbook overfitting.
  • Lesson: ML on features with no a-priori economic thesis = noise-fit.

SentimentHybrid.py

Section titled “SentimentHybrid.py”
  • Why retired: Earlier hybrid attempt. Never passed OOS.
  • Lesson: you can’t validate an ML model on the same data you trained it on.

SentimentRL.py / SentimentRL1h.py

Section titled “SentimentRL.py / SentimentRL1h.py”
  • Why retired: PPO/A2C reinforcement learning agents. Learned to “buy-and-hold with extra steps” in-sample. OOS: total loss, then gave up.
  • Lesson: RL on small financial samples (few million ticks) overfits catastrophically. You’d need millions of episodes across many markets.

SentimentFreqAI.py / SentimentFreqAIClassifier.py

Section titled “SentimentFreqAI.py / SentimentFreqAIClassifier.py”
  • Why retired: FreqAI pipelines with LightGBM, XGBoost, PyTorch MLP, Transformer. All produced beautiful in-sample fits (Calmar > 50) that vanished out-of-sample.
  • Lesson: ML models trained on n < 5000 examples with 50+ features will memorize, not generalize. Guaranteed.

Early Iterations (superseded)

Section titled “Early Iterations (superseded)”

TrendFollowEMA.py

Section titled “TrendFollowEMA.py”
  • Why retired: Earliest trend-follower. Replaced by HonestTrendGeneric and its subclasses. Functionally similar but without FnG defense, risk-manager integration, or walk-forward validation.
  • Lesson: the first version of any strategy is a prototype, not a product.

DonchianBreakout.py

Section titled “DonchianBreakout.py”
  • Why retired: Breakout attempts on daily timeframe. Too few trades/year (~8) to validate. OOS inconclusive.
  • Lesson: low trade frequency makes any backtest statistically useless regardless of apparent results.

Dead Edges (looked real, weren’t)

Section titled “Dead Edges (looked real, weren’t)”

PairMeanReversion.py

Section titled “PairMeanReversion.py”
  • What it was: ETH/BTC z-score mean reversion. Buy ETH when z < −2 (ETH undervalued vs BTC), exit when z reverts to 0.
  • The seductive stat: p < 0.0001 in-sample (n=2382, avg +0.96% per week).
  • What happened: 2025 saw persistent ETH/BTC trend drift. The mean- reversion assumption broke. OOS: lost money.
  • Lesson: statistical significance on historical data does not guarantee the underlying regime persists. “Every strategy dies cliff-fall” (Taleb). Always have a kill-switch on regime change.

HonestTrend Earlier Variants (replaced by Generic + Subclasses)

Section titled “HonestTrend Earlier Variants (replaced by Generic + Subclasses)”

These predate the HonestTrendGeneric + risk-manager architecture. They were rungs on the ladder, not the ladder itself.

HonestTrend.py (daily)

Section titled “HonestTrend.py (daily)”
  • Why retired: ~10 trades/year. Can’t validate anything with n=20 across 2 years.
  • Lesson: daily-timeframe trend following on 3 pairs gives insufficient sample for significance tests.

HonestTrend1h.py (1h)

Section titled “HonestTrend1h.py (1h)”
  • Why retired: OOS −14% to −25% across multiple parameter choices. Stage 2 cross-timeframe test later confirmed: 1h timeframe completely fails for this edge, regardless of EMA choice.
  • Lesson: “signal works at X timeframe” is often shorthand for “execution works at X timeframe”. Latency kills.

HonestTrend1m.py (1m, original)

Section titled “HonestTrend1m.py (1m, original)”
  • Why retired: Superseded by HonestTrend1mLive.py subclass. Kept here as the original before the risk-manager framework.
  • Lesson: add risk-management before you deploy, not after.

HonestTrend15m.py + .json (15m, original)

Section titled “HonestTrend15m.py + .json (15m, original)”
  • Why retired: Superseded by HonestTrend15mDry.py. This was the hyperopt-fitted version (EMA 94/139, ratio 1.48x) before we understood whether the fit was real.
  • Lesson: hyperopt results must be walk-forward validated before being treated as “the” parameters.

HonestTrend15m_2xRatio.py + .json

Section titled “HonestTrend15m_2xRatio.py + .json”
  • What it was: Stage 1 testing artifact. Fixed slow = 2×fast, varied fast across {24, 48, 72, 96, 128, 160}.
  • Why retired: Served its purpose — proved the edge is real (non- hyperopt 2x ratios also profit OOS) but not as robust as 1.48x ratio.
  • Lesson: keep testing scaffolding separate from production code.

HonestTrend4h.py (4h)

Section titled “HonestTrend4h.py (4h)”
  • Why retired: Best PF (1.74 OOS) but only ~15 trades/year. Too few for meaningful validation. Not a bad strategy; just not enough data to trust.
  • Lesson: a 1.74 PF over 15 trades means your 95% CI on PF is roughly [0.7, 4.5]. Useless.

Restoration Protocol

Section titled “Restoration Protocol”

If a future you wants to revive one of these:

  1. Read the “Why retired” entry above first. The failure mode is documented. Don’t rediscover it.
  2. Run the full validation stack: Stage 1 (parameter stability) + Stage 2 (cross-timeframe consistency) + Stage 3 (fat-tail + bootstrap significance) from HONEST_TREND_REPORT.md.
  3. Integrate risk_manager.py before any dry-run.
  4. Parallel dry-run at least 3 months before risking real money.
  5. Do not trust in-sample metrics. Only walk-forward.

What To Read When Things Go Wrong

Section titled “What To Read When Things Go Wrong”
  • HONEST_TREND_REPORT.md — full decision record
  • ../strategies/risk_manager.py — kill-switch rules
  • ../risk_state.json — live state
  • ../walk_forward_history/ — monthly re-validation archive

If you’re reading this because the live strategy died, the right reaction is not “try another strategy from legacy”. The right reaction is: is the entire class of strategy (short-term trend following) no longer working? Or did a specific implementation break? Answer that first, then decide whether to build new or restart from legacy.