Class GridSearchOptimizer

java.lang.Object

com.quantfinlib.backtest.validation.GridSearchOptimizer

public final class GridSearchOptimizer
extends Object

Exhaustive strategy parameter search: backtests every grid combination and ranks by an objective (e.g. PerformanceMetrics::sharpeRatio). Non-finite objectives rank last. Feed the resulting in-sample winners into WalkForwardAnalyzer — never trust them raw.

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static final class	GridSearchOptimizer.Candidate	One grid point: its parameters, metrics, objective and equity curve.

Method Summary

Modifier and Type	Method	Description
static GridSearchOptimizer.Candidate	best(ParameterGrid grid, StrategyFactory factory, BarSeries series, BacktestConfig config, ToDoubleFunction<PerformanceMetrics> objective)	The winning parameter set only.
static double	deflatedSharpeOfWinner(List<GridSearchOptimizer.Candidate> ranked, double[] winnerReturns, int periodsPerYear)	The multiple-testing haircut for the grid's winner: the probability that the top-ranked candidate's Sharpe beats what the best of effectiveTrials(List) zero-skill trials would have scored anyway (the deflated Sharpe of SharpeValidation.probabilisticSharpe(double, double, int, double, double) against E[max] with N̂ in place of the raw grid size).
static double	effectiveTrials(List<GridSearchOptimizer.Candidate> ranked)	Effective number of independent trials in a search, N̂ = ρ̄ + (1 − ρ̄)·M for M trials whose per-bar return streams have mean pairwise correlation ρ̄ (clamped to [0, 1]); the average-correlation approximation in Bailey and López de Prado (2014, J.
static List<GridSearchOptimizer.Candidate>	search(ParameterGrid grid, StrategyFactory factory, BarSeries series, BacktestConfig config, ToDoubleFunction<PerformanceMetrics> objective)

Methods inherited from class Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Details

search

public static List<GridSearchOptimizer.Candidate> search(ParameterGrid grid,
 StrategyFactory factory,
 BarSeries series,
 BacktestConfig config,
 ToDoubleFunction<PerformanceMetrics> objective)

best

public static GridSearchOptimizer.Candidate best(ParameterGrid grid,
 StrategyFactory factory,
 BarSeries series,
 BacktestConfig config,
 ToDoubleFunction<PerformanceMetrics> objective)

The winning parameter set only.

effectiveTrials

public static double effectiveTrials(List<GridSearchOptimizer.Candidate> ranked)

Effective number of independent trials in a search, N̂ = ρ̄ + (1 − ρ̄)·M for M trials whose per-bar return streams have mean pairwise correlation ρ̄ (clamped to [0, 1]); the average-correlation approximation in Bailey and López de Prado (2014, J. Portfolio Management 40(5), §3, following the effective number of tests of Bailey, Borwein, López de Prado and Zhu). Adjacent grid points (fast 10/slow 50 vs fast 10/slow 55) trade almost the same book; counting them as independent overstates the expected maximum.

deflatedSharpeOfWinner

public static double deflatedSharpeOfWinner(List<GridSearchOptimizer.Candidate> ranked,
 double[] winnerReturns,
 int periodsPerYear)

The multiple-testing haircut for the grid's winner: the probability that the top-ranked candidate's Sharpe beats what the best of effectiveTrials(List) zero-skill trials would have scored anyway (the deflated Sharpe of SharpeValidation.probabilisticSharpe(double, double, int, double, double) against E[max] with N̂ in place of the raw grid size). A grid search computes a Sharpe for every trial and then quietly reports only the maximum — this is the number that accounts for that selection. Values near 1 mean the winner survives its own search; below ~0.95 the "best" parameter set is indistinguishable from picking the luckiest of N random ones.

Parameters:

ranked - result of search(ParameterGrid, StrategyFactory, BarSeries, BacktestConfig, ToDoubleFunction) (uses every trial's Sharpe as the null distribution), ≥ 2 trials

winnerReturns - the winner's per-period returns (derive from its equity curve), ≥ 4 observations

periodsPerYear - the annualization used by the backtest metrics