Hyperopt Pyramid tuning

Hyperopt Experiment: Pyramid Parameter Optimization

Goal: Hand the 4 pyramid parameters I set “by feel” over to optuna, use a custom loss to avoid overfitting, and walk-forward validate whether it is actually better.

Methodology

1. Parameter Space

pyramid_1_trigger     = DecimalParameter(0.03, 0.10)  # original default 0.05
pyramid_2_trigger     = DecimalParameter(0.08, 0.20)  # original default 0.12
pyramid_1_stake_ratio = DecimalParameter(0.30, 1.00)  # original default 0.60
pyramid_2_stake_ratio = DecimalParameter(0.20, 0.80)  # original default 0.40

2. Custom Loss — HonestHyperOptLoss

Avoid the pitfalls of built-in losses:

• ShortTradeDurHyperOptLoss encourages short-term trading (violates our mid-term philosophy)
• SharpeHyperOptLoss is easily fooled by big winners
• Built-in CalmarHyperOptLoss is closest, but has no hard DD threshold

Our version (strategies/HonestHyperOptLoss.py):

IF trade_count < 50:       loss = 999  (reject low sample size)
IF max_dd > 20%:           loss = 100 + max_dd * 10  (hard reject)
IF max_dd > 15%:           Calmar divided by penalty factor (1.0 → 2.0)
ELSE:                      loss = -Calmar  (higher Calmar is better)

3. Training / Validation Split

• Training: 2022-01 → 2024-12 (3 years, 3 regimes: LUNA crash / recovery / ETF bull)

• Validation:

  • True out-of-sample past: 2018-2020 (W1-W3)
  • True out-of-sample future: 2025-present (W8)
  • Contained in-sample: 2022-2024 (W5-W7)

• CPU: Single-threaded (multi-threading could not be enabled due to joblib pickle errors)

• Optuna sampler: NSGAIIISampler (default, multi-objective friendly)

• Epochs: 100

• Runtime: 19 minutes

Results

Best Parameters (epoch 42/100)

Parameter	Original default	Hyperopt best	Change
pyramid_1_trigger	0.05	0.08	+60%
pyramid_2_trigger	0.12	0.10	−17%
pyramid_1_stake_ratio	0.60	0.80	+33%
pyramid_2_stake_ratio	0.40	0.80	+100%

Interpretation:

• First add-on threshold raised to +8% (more conservative, filters noise)
• Second add-on threshold instead dropped to +10% (once the first is cleared, follow up faster)
• Both add-on sizes expanded to 0.8× initial (max total position reaches 2.6× initial vs the old 2.0×)
• Essence: more aggressive pyramid, but stricter thresholds

Training Set Performance (2022-01 → 2024-12)

Trades: 206
Win Rate: 37.4%
Total Profit: +48.86%
Max DD: 10.31%
Calmar Objective: -4.74

Full-History Comparison (2017-08 → 2026-04)

Metric	Old default	Hyperopted	Difference
Total Profit	+195%	+207%	+12 ppt
Trades	570	570	0
Win Rate	33.2%	35.3%	+2.1 ppt
Max DD	16.93%	15.53%	−1.4 ppt
Avg Trade	0.15%	−0.08%	Slightly lower (trades higher win rate for higher profit mode)

8-Regime Walk-Forward Comparison

Window	Old default	Hyperopted	Difference	Notes
W1 2018 crash	−14.43%	−13.27%	+1.16	🎯 OOS improvement
W2 2019 accum	+9.57%	+8.42%	−1.15	OOS slightly lower
W3 2020 COVID	+46.88%	+45.74%	−1.14	OOS slightly lower
W4 2021 bull	+34.77%	+31.54%	−3.23	Pre-training
W5 2022 LUNA	−4.83%	−4.68%	+0.15	Training set
W6 2023 recovery	+4.72%	+5.90%	+1.18	Training set
W7 2024 ETF	+44.69%	+47.29%	+2.60	Training set
W8 2025 present	−4.00%	−2.46%	+1.54	🎯 OOS improvement
			+1.11 sum
profitable/8	5/8	5/8	0	Stable

Key observations:

1. Training set (W5-W7) cumulative +3.93 ppt improvement — as expected
2. OOS past (W1-W3) cumulative −1.13 ppt — slight degradation on past data, but W1 (hardest 2018 window) actually improved +1.16
3. OOS future (W8) +1.54 ppt improvement — beneficial for the current market
4. No overfitting signal: OOS averages +0.24 ppt improvement, beyond 1/8 sample significance

Conclusion: Worth Adopting

✅ Pros:

• Full-history +12 ppt return + DD down 1.4 ppt
• Both W1 2018 (hardest window) and W8 2025 (current) improved
• No overfitting evidence (OOS data shows net positive improvement)

⚠️ Cons:

• W4 2021 bull-market top −3.23 ppt (marginal degradation in high-valuation era)
• W2/W3 OOS slight degradation (still profitable)

Already Applied

1. strategies/HonestTrendGeneric.py base class defaults updated to hyperopt-best values
   • All subclasses (HonestTrend15mDry, 1mLive, 1mMTF) inherit automatically
2. strategies/HonestTrend15mDry.json auto-saved by freqtrade (takes precedence for that class specifically)
3. Production configs already have pyramid enabled (see docs/EXPERIMENTS_DCA_AND_PYRAMID.md)

Extended Experiment: 6-Parameter Hyperopt (❌ Rejected)

2026-04-21 follow-up test: Added adx_threshold and min_hold_minutes to the hyperopt space (6 parameters total), 150 epochs.

Results

Training set performance was off the charts:

Objective: -10.84  (vs -4.74 in 4-param baseline)
In-sample (2022-2024): 119 trades, 37.8% WR, +70%, Max DD 6.46%

Best parameters:

• adx_threshold: 18 → 30 (sharply raised entry threshold)
• min_hold_minutes: 720 → 1065 (~18h minimum hold)
• pyramid_1_stake_ratio: 0.80 → 1.00 (first add-on maxed out)

Walk-forward 8 windows: 6/8 (looks better!)

Window	4-param	6-param	Diff
W1 2018	−13.27%	−20.02%	−6.75 ❌
W4 2021	+31.54%	+58.59%	+27.05 ✅
W5 2022	−4.83%	+4.89%	+9.57 ✅ flipped to positive
W7 2024	+47.29%	+55.02%	+7.73 ✅
Sum			+19.4 ppt

❌ But Full History Is a Disaster

Metric	4-param	6-param	Diff
Total profit	+207%	+160%	−47 ppt
Max DD	15.53%	24.81%	+9.3 ppt 🚨
Trades	570	335	−40%

24.81% max DD exceeds the 20% kill-switch threshold → would trigger auto-RETIRE in live trading.

Why Good on Walk-Forward but Bad on Full History?

• Walk-forward each window uses an independent $10K principal, peak-to-trough only within the year
• Full history compounds, and the −6.7 ppt loss in mid-2018 stacks on top of already-accumulated bull-market profits = bigger percentage drawdown
• ADX=30 strict filtering rallies harder in bull markets, but recovers to new highs later in bear markets due to fewer entries
• Simple “profitable each window?” ignores the relative depth during drawdowns

Rollback Decision

• adx_threshold and min_hold_minutes restored to optimize=False
• strategies/HonestTrend15mDry.json restored to the old best values from 4-param hyperopt
• Lesson: walk-forward profitable rate ≠ full-history max DD control

Extended Experiment: CmaEsSampler (Technical Improvement)

Added a generate_estimator method (inside the nested HyperOpt class) that uses CmaEsSampler to replace the default NSGAIIISampler. CmaEs converges faster on continuous parameters.

Auto-enabled for future hyperopts, but the rejected 6-param experiment used NSGAIII (before switching to CmaEs), so the results are unaffected.

Extended Experiment: HonestTrend1mLive Standalone Hyperopt (❌ Rejected)

2026-04-22: Standalone hyperopt on the 1m timeframe (base class was tuned on 15m data; in theory, 1m might have different optima).

Setup

• Training range: 2024-01 → 2024-12 (1 year 1m BTC+ETH)
• Epochs: 50, min_trades=25 (fewer trades on 1m, must lower the threshold)
• Space: 4 pyramid parameters (leave EMA/ADX/min_hold alone)
• Sampler: CmaEs

Best Parameters (epoch 30/50)

Parameter	base class default	1m hyperopt	Change
pyramid_1_stake_ratio	0.80	0.60	−25%
pyramid_2_stake_ratio	0.80	0.70	−13%
pyramid_2_trigger	0.10	0.08	−20%
pyramid_1_trigger	0.08	0.08	=

Direction: more conservative add-ons (smaller stake, lower threshold but smaller amount on the second add-on).

Training Set (2024 single year)

• Default: +25.21%
• 1m hyperopt: +24.72% (−0.49 ppt)

Already no clear improvement on the training set — indicating CmaEs cannot find a better combo than the default within 50 epochs.

Full-History 3.3-Year Comparison (2023-01 → 2026-04)

Metric	Base class default	1m hyperopt	Diff
Total Profit	+37.99%	+36.13%	−1.86 ppt
Max DD	10.55%	10.33%	−0.22 ppt
Calmar (wallet)	4.15	4.04	−0.11
Trades	135	135	0
Win Rate	35.6%	35.6%	0

Rollback Decision

• Delete strategies/HonestTrend1mLive.json
• HonestTrend1mLive continues to inherit base class pyramid defaults
• Lesson: the same 4 pyramid parameters are good enough on both 15m and 1m; standalone hyperopt on 1m is noise-chasing

Why No Improvement?

1. Pyramid triggers are “profit percentage”, independent of timeframe — +8% is +8% on both 15m and 1m
2. What actually varies with timeframe is EMA/ADX, but we explicitly do not hyperopt those
3. 50 epochs + 1 year data is not enough for CmaEs to converge to any meaningful improvement

Future Optimization Roadmap

Short-Term

• Observe 1-2 months of live W8 (2025-present) to confirm the 4-param new parameters are actually better

Don’t Do

• Do not add adx/min_hold to the hyperopt space again (verified above to be a trap)
• Do not tune >5 parameters at once (overfitting probability grows exponentially with dimensionality)
• Do not put EMA periods in hyperopt (these are “structural” choices, not something the optimizer should randomly change)
• Do not hyperopt 1m vs 15m pyramid parameters separately (verified as noise-chasing on 2026-04-22)

Key Lessons

1. Lower training loss ≠ better production: objective going from -4.74 → -10.84 looks like 2× improvement on the training set, but full-history DD multiplied by 1.6×
2. Walk-forward independent windows ≠ full-history backtest: the former misses the compounding damage of “continuous disadvantage”
3. Trust the kill-switch threshold: 20% DD was not set arbitrarily — it is the psychological upper bound the live account can withstand. Letting hyperopt push DD to that threshold is playing with fire
4. Less is more: 4 pyramid parameters + strict loss is enough. Adding 2 more variables just injects noise
5. Pyramid parameters are timeframe-agnostic across timeframes: the same 4 pyramid parameters are optimal on both 15m and 1m — because triggers look at percentage profit, not candle count

Reproduction

# Hyperopt
freqtrade hyperopt \
  --strategy HonestTrend15mDry \
  --config configs/backtest/config_backtest_15m_btceth.json \
  --datadir user_data/data/binance \
  --user-data-dir user_data \
  --strategy-path strategies \
  --hyperopt-loss HonestHyperOptLoss \
  --hyperopt-path strategies \
  --spaces buy \
  --epochs 100 \
  --timerange 20220101-20241231 \
  -j 1  # single-threaded due to joblib pickle issue with our strategy

# Inspect best epoch
freqtrade hyperopt-show --best \
  --user-data-dir user_data \
  --hyperopt-filename strategy_HonestTrend15mDry_2026-04-21_17-08-21.fthypt

# Walk-forward validation
python scripts/walk_forward_full_history.py \
  --strategy HonestTrend15mDry \
  --timeframe 15m \
  --config configs/backtest/config_backtest_15m_btceth.json

References

• Freqtrade Advanced Hyperopt
• WALK_FORWARD_FULL_HISTORY.md — 8-regime methodology
• TUTORIAL_FOR_BEGINNERS.md §1.8 — Hyperopt pitfalls