The FX Parameter Sweep That Didn't Find a Needle in the Haystack¶

Introduction¶

Parameter sweeps are a fundamental tool in quantitative trading: you vary a key parameter across a range and look for a sweet spot—the point where performance metrics peak and the strategy becomes profitable.

The promise is seductive: maybe your strategy failed at a sub-optimal parameter setting, and a systematic grid search will surface a hidden configuration that works.

But sometimes, the grid sweep reveals the opposite: a global optimum that is still worse than the acceptance gate. This is the story of a Bollinger Reversion strategy tested across three major FX pairs, where the best-in-class result was still a Sharpe ratio of -0.33, missing the > 1.0 viability threshold by a wide margin.

The Setup: Revisiting a Rejected Strategy¶

The Bollinger Reversion strategy (mean-reversion on Bollinger Band extremes) had been tested previously on a single pair and rejected. The hypothesis behind this new sweep was reasonable: what if the original rejection point was at a sub-optimal Bollinger Reversion parameter?

The revised plan: 1. Run a full parameter grid sweep of the bb_k parameter (the number of standard deviations from the mean) across 2.6 to 3.6 in 0.1 increments—33 combinations 2. Test across three major pairs: GBPUSD, USDCAD, AUDUSD 3. Apply realistic IG (Tradable Instrument) spread modeling (intraday spreads: 1–3 pips depending on pair and time) 4. Include 8-year walk-forward validation for any configuration crossing the > 1.0 Sharpe gate 5. Check correlation against existing deployed strategies (target: < 0.4 average correlation)

The expectation: at least one pair + parameter combination would reach profitability.

The Results: No Needle Found¶

After the full 33-cell grid across 3 FX pairs:

Global verdict: CONFIRMED KILL

Not a single parameter-pair combination cleared the viability gate of Sharpe > 1.0. Not even close. The best result (GBPUSD at bb_k=2.8) was negative, with a profit factor under 1.0 (losing money per round trip on average).

Why Parameter Sweeps Can Mislead¶

The False Premise: "Sub-Optimal Parameter"¶

The underlying assumption of a parameter sweep is that you're searching within the problem space of a viable strategy. That is: the strategy's architecture is sound, but you haven't found the right configuration yet.

This assumption often fails because the problem isn't the parameter—it's the core logic.

Bollinger Reversion's Fundamental Mismatch¶

Mean-reversion strategies assume that price extremes tend to revert toward the mean. This is true in some regimes (highly mean-reverting, low-trend environments) and false in others (trending, momentum-driven environments).

FX pairs in 2020–2026 have been characterized by: - Dollar strength episodes: Structural USD appreciation trends lasted months (2021, 2022, 2024), creating extended mean-shift conditions where Bollinger Reversion sold repeatedly into a continuing trend - Pairs with directional bias: AUDUSD reflected commodity cycle trends; GBPUSD reflected Brexit and UK rate regime shifts. Neither was mean-reverting at the timeframe tested - Momentum persistence: Post-COVID, FX volatility expanded and created stronger trend-following opportunities—the opposite of mean reversion

A parameter sweep can't fix a strategy that's working against the regime's dominant behavior. Tightening or loosening the Bollinger bands (what bb_k controls) just changes how quickly you lose money, not whether you lose it.

The Spread Bleed on Multiple Exits¶

Bollinger Reversion implies frequent exits: - Entry on band extremes (multiple times per day on volatile pairs) - Exit on mean-crossing (rapid profit-taking or stop-loss at the midband) - Result: ~20–40 round trips per pair over a testing window

Each round trip on FX incurs spread costs: - GBPUSD: ~1.2 pips typical (£0.12 per £10k position) - USDCAD: ~1.5 pips - AUDUSD: ~1.5–2.0 pips (lower liquidity)

Over 20+ round trips, the cumulative spread bleed often exceeds the intended alpha by 50–200%. The parameter sweep didn't address this arithmetic—it just confirmed it.

Deeper Learning: The Problem with Optimism Bias in Grid Sweeps¶

Parameter sweeps are prone to a subtle form of overfitting: optimism bias in result interpretation.

When you run a 33-cell grid and the best cell is "only slightly negative," there's a temptation to think: "Well, maybe with slightly better execution or a slightly different market regime, this could work."

The data this time was unambiguous: even the best-case configuration lost money. There was no debate about whether the strategy was borderline—it had failed decisively across a full grid under realistic conditions.

Implications for Strategy Development¶

This sweep illustrates several principles:

1. Parameter Sweeps Test Configuration, Not Viability¶

A grid sweep is useful for fine-tuning a strategy that's already in the ballpark of profitability. If you're starting from a negative Sharpe, a sweep is likely to yield no surprises—different points on a losing curve.

2. Regime Matters More Than Parameters¶

Before running a sweep, ask: "Is this strategy architecturally aligned with the current market regime?" - Bollinger Reversion works best in low-trend, mean-reverting regimes - FX 2020–2026 exhibited strong directional biases - The parameter sweep was addressing the wrong problem

3. Frequency and Spread Economics Are First-Order¶

A strategy generating 30+ round trips per instrument is inherently dependent on spread efficiency. If spread costs exceed 50% of alpha, no parameter tweak will rescue you. Test spread assumptions before running a grid.

Conclusion¶

This parameter sweep delivered a clear but humbling lesson: when a grid reveals no viable configuration, it's usually because the strategy's premise doesn't fit the environment, not because the parameters are wrong.

Bollinger Reversion on these FX pairs during this period was fighting against the market regime's dominant characteristics (trend, momentum). A parameter sweep that searches {2.6, 2.7, 2.8, ... 3.6} can't overcome that mismatch.

The efficiency of this outcome, though, is valuable: we tested 99 configurations (33 parameters × 3 pairs) and learned conclusively that none worked. This eliminated a hypothesis from further consideration, freeing resources for strategies better suited to the actual market environment.

Sometimes the best result of a parameter sweep is a decisive kill decision—not a profitable configuration.

Data sources: FX pair backtest data (GBPUSD, USDCAD, AUDUSD, 2020–2026), spread modeling from IG Tradable Instruments data.

Note on methodology: This analysis is based on historical backtest data for major currency pairs using standard Bollinger Band indicators and mean-reversion entry logic. No proprietary execution models or live trading data are referenced.