DeTime Method-Specific Strategy Expansion

This tutorial layer extends the two core strategy families already implemented in the strategy lab:

1. Trend following: trade when the decomposed trend is strong enough to define a directional regime.
2. Oscillation reversion: trade when the decomposed trend is weak and the residual becomes the tradable deviation.

The new point is that the decomposition method and decomposition horizon are not neutral implementation details. They create different trend, cycle, and residual estimates, and therefore create different strategies.

1. Why different decomposition methods create different strategies

A raw moving-average strategy changes when the moving-average length changes. A decomposition strategy changes when any of these choices change:

• the decomposition method: STL, SSA, STD, Wavelet, EMD, VMD, etc.;
• the decomposition period: short cycle, medium cycle, long cycle;
• the walk-forward training window;
• the recomputation step;
• the residual z-score window;
• volume decomposition and liquidity filters.

The same trading rule can be held fixed:

enter long if trend_slope > 0 and trend_strength is high
enter reversion if residual_z is too negative in a weak-trend regime

but if STL and SSA extract different trends and cycles, then the entry dates, exits, position sizes, and backtest results are different. The tutorial therefore treats each method-horizon combination as a separate strategy variant.

2. Method selection guide

Method	Trading interpretation	Useful default	Main risk
STL	fixed-period seasonal-trend split	stable daily/weekly/monthly rhythm	phase errors when the cycle drifts
SSA	low-rank Hankel subspace split	smoother quasi-periodic financial cycles	leakage under abrupt breaks
STD	seasonal-trend-dispersion split	when volatility/dispersion matters	may be conservative on entry signals
Wavelet	multi-scale local split	local regime changes and multi-scale cycles	requires PyWavelets and scale tuning
EMD/CEEMDAN	adaptive intrinsic oscillatory modes	non-stationary amplitude/frequency cycles	mode mixing
VMD	narrow-band spectral modes	compact frequency bands	sensitive to K and alpha

3. Parameter tuning logic

The tutorial avoids blind optimization first. It uses interpretable grids:

Market idea	DeTime parameter	Example
medium swing rhythm	period	63 bars
half-year structure	period	126 bars
annual structure	period	252 bars
long trend structure	train_window	504 bars
fast adaptation	step	5 or 21 bars
low turnover	step	21 or 63 bars
residual band sensitivity	z_window	63 or 126 bars

A practical workflow is:

1. Start with STL, SSA, and STD on periods 63, 126, and 252.
2. Compare strategy families, not only decomposition plots.
3. Check orders, round-trip trades, drawdown, turnover, and exposure.
4. Add Wavelet only after installing PyWavelets.
5. Treat the best variant as a hypothesis, not as proof of alpha.

4. Strategy variants now implemented

The module examples/quant_trading/strategy_method_variants.py implements:

• decomposition_trend_following_signals
• decomposition_oscillation_reversion_signals
• decomposition_residual_bollinger_signals
• decomposition_macd_trend_signals
• decomposition_trend_crossover_signals
• run_method_variant_grid

The residual Bollinger strategy replaces raw-price bands with residual bands around trend + cycle fair value. The trend MACD strategy computes MACD on the decomposed trend instead of raw price. The trend crossover strategy computes fast/slow EMAs on the decomposed trend.

5. Component-level pair trading

The module examples/quant_trading/strategy_component_pairs.py implements decomposition-first pair trading.

The logic is:

1. Decompose both assets.
2. Check whether their trends are similar.
3. Check whether their cycles are similar.
4. Use Engle-Granger and ADF diagnostics on raw price, fair value, and spread residuals.
5. Trade the residual gap or the deviation from the trend+cycle relationship.

Implemented pair strategies:

• classic_pair_spread_zscore
• detime_<method>_component_residual_gap
• detime_<method>_fair_spread_deviation
• detime_<method>_cointegration_filtered_residual_gap

The pair-trading tutorial is not only about whether two raw prices are cointegrated. It also asks whether the decomposed components support the pair hypothesis:

trend_left ~ trend_right
cycle_left ~ cycle_right
residual_left - residual_right is the traded gap

6. Commands

Offline real-data smoke run:

make strategy-expansion

Live Yahoo Finance run:

make strategy-expansion-live

Direct custom run:

python examples/quant_trading/scripts/run_strategy_expansion.py \
  --use-bundled-sample \
  --variant-grid custom \
  --methods STL SSA STD \
  --periods 63 126 252 \
  --train-window 504 \
  --step 21 \
  --pair-method STL \
  --pair-period 126

7. Output files

The script writes:

examples/quant_trading/reports/strategy_expansion/method_variant_strategy_stats.csv
examples/quant_trading/reports/strategy_expansion/method_variant_orders.csv
examples/quant_trading/reports/strategy_expansion/method_variant_trades.csv
examples/quant_trading/reports/strategy_expansion/method_variant_spec_grid.csv
examples/quant_trading/reports/strategy_expansion/method_variant_feature_coverage.csv
examples/quant_trading/reports/strategy_expansion/method_variant_failed_methods.csv
examples/quant_trading/reports/strategy_expansion/component_pair_strategy_stats.csv
examples/quant_trading/reports/strategy_expansion/component_pair_diagnostics.csv
examples/quant_trading/reports/strategy_expansion/component_pair_orders.csv
examples/quant_trading/reports/strategy_expansion/component_pair_trades.csv
examples/quant_trading/reports/strategy_expansion/strategy_expansion_run_manifest.json

Smoke outputs are tutorial evidence that the pipeline runs. They are not production trading claims.