Dispersion and Relative Strength Backtests Need Proxy Discipline

Abstract

Relative-value intraday strategies depend on aligned context. If a strategy compares QQQ against SPY, or one ETF against a sector proxy, the backtest has to align those series bar by bar and avoid letting the proxy introduce hidden lookahead.

For developers, dispersion research is useful because it forces the system to handle multiple streams honestly.

What The Proxy Does

A proxy series can define beta-adjusted spread, relative strength, shock filters, or confirmation. The traded symbol is not evaluated in isolation. It is evaluated against what a related market was doing at the same time.

That can improve a signal, but it also adds failure modes. Missing proxy bars, mismatched timestamps, and unstable beta estimates can all create false precision.

Define The Alignment Contract

Proxy research should start with an alignment contract. The contract states which bar timestamps are eligible, how missing bars are handled, whether bars must be present for both series, and when rolling statistics are computed. Without this contract, the replay engine can accidentally compare a fresh primary bar with a stale proxy bar.

A strict default is to require both series to have completed bars for the same timestamp before scoring the signal. If one side is missing, the event rejects or carries a documented missing-data status. Forward filling can be useful for some macro features, but it is risky for intraday relative-strength signals because the stale side can make the spread look more extreme than it was.

The contract should also define session boundaries. A proxy based on regular-session bars should not mix premarket context unless the strategy explicitly uses it. These details sound small, but they determine whether the comparison is measuring relative behavior or timestamp mismatch.

Align Before You Score

The replay engine should align primary and proxy bars before computing spread or z-score. It should also define the rolling window used for beta or correlation. If the correlation filter uses future bars, the signal becomes contaminated.

The same next-bar discipline still applies. A relative-strength signal generated from completed bars should enter only after those bars were observable.

Estimate Beta Causally

Beta-adjusted spread signals are only as clean as the beta estimate. If the estimate uses future returns, the signal has already seen part of the answer. A causal beta estimate should use only data available before the signal timestamp, and the lookback window should be logged as part of the experiment.

The estimate should also be stress-tested. A short lookback can adapt quickly but become noisy. A long lookback can be stable but slow to reflect regime changes. If a dispersion branch only works for one narrow beta lookback, the result may be a parameter artifact rather than a robust relative-value observation.

Developers can add a simple diagnostic: compare raw spread, beta-adjusted spread, and rank-based relative strength. If all three point to similar events, the signal has stronger structural support. If only one transformation works, the next step is to understand why before promoting the branch.

Why Direction Matters

Dispersion signals are often not symmetric. A branch that works when QQQ leads SPY may fail when the same idea is inverted or moved to another symbol. Developers should not assume a relative-value family is portable just because the math is reusable.

This is where no-go reports help. They show whether a failure came from signal logic, proxy mismatch, option expression, or portfolio overlap.

Use Null Tests For Proxy Contribution

A proxy should earn its place in the model. One way to test that is to run null variants: primary-only signal, proxy-only filter, shuffled proxy timestamps, or a related but less appropriate proxy. The goal is to see whether the proxy adds information or only adds another parameter surface.

If the primary-only version performs similarly, the dispersion layer may be unnecessary. If a shuffled proxy performs well, the apparent edge may be coming from general market regime rather than aligned relative strength. If a less appropriate proxy performs better, the original hypothesis needs revision.

These tests are not meant to be punitive. They protect the research from elegant but unhelpful complexity. A proxy should improve timing, reduce bad trades, or clarify regimes. If it does none of those, it may be adding implementation risk without adding evidence.

Connect The Proxy Signal To The Option Expression

The relative-strength signal may live in the underlying pair, but the trade still has to be expressed through options. That creates a second layer of asymmetry. A proxy event that looks clean on the underlying can select contracts with different liquidity, DTE availability, and spread behavior depending on direction.

For example, a long-call expression after a relative-strength breakout can face a different quote surface than a long-put expression after a relative weakness event. The signal math may be symmetric while the option market is not. A serious report should therefore show rejects and fill quality by direction, not only combined PnL.

This also matters for portfolio construction. A dispersion branch can look useful alone but overlap heavily with broad index momentum or volatility sleeves. The proxy signal should be evaluated for independent contribution, not only standalone return.

When A Proxy Branch Should Stay A Research Note

A proxy-based strategy should stay in research when alignment is fragile, beta estimates are unstable, direction changes reverse the conclusion, or option rejects cluster around the exact events that generate the signal. These outcomes do not mean the idea is worthless. They mean the system has not yet converted the relative-value observation into a tradable rule.

That distinction is important for developers. Closing a branch is not the same as discarding the learning. A no-go report can preserve the alignment assumptions, proxy diagnostics, and option-expression failures so the next experiment starts from evidence rather than memory.

Takeaway

Proxy-based research is powerful only when the data alignment is boring and strict. Build the alignment first, then test the strategy. Otherwise the proxy can become a hidden source of hindsight.