Dispersion Trading Backtest: QQQ vs SPY and Why the Signal Was Not Symmetric

Repository reference: cutebacktests

Abstract

Dispersion sounds like one theme, but in practice the edge can be concentrated in one sleeve and absent in another. This repository's recent work on relative-value and dispersion-style intraday breakouts makes that unusually clear. The strongest tactical insight from the dispersion branch was not "dispersion works." It was "QQQ carried much more of the signal than SPY."

The key evidence appears in Episode 9 and Toward The One Piece Of Sharpe. The strict-causal reruns showed QQQ single +20.81% on 5 trades and QQQ vertical +10.85% on 8 trades, while the best positive SPY paths remained weak or extremely sparse. Later Stage B replays reported qqq_single_base with 9 trades and +44537.92, and qqq_single_volume15_v1 with 10 trades and +32998.35. Those are strong-looking numbers, but the repo still kept the QQQ sleeve as research_only because the sample remained thin.

Question

The practical question is not whether dispersion is a good idea in the abstract. The real question is where the signal actually lived once the branch was decomposed by index sleeve, overlay, and parity quality.

That matters because combined dispersion results can hide a very uneven picture. If one sleeve is carrying the economic value and the other is weak or negative, the combined result can be harder to interpret than either component on its own. This repo learned that symbol decomposition was not a side exercise. It was the main way to understand the branch.

Method: How the Dispersion Trading Backtest Was Decomposed

As Episode 9 explains, the QQQ sleeve is not a totally separate family. It is a specialized descendant of the c4 dispersion-relative breakout logic. At the stock level, it uses QQQ-specific variants of the dispersion breakout seed, adjusting relative-volume floors, beta-shock tolerance, and entry thresholds. At the option level, the repo later moved away from fixed overlays and derived the option overlay from the actual stock winner selected upstream.

This decomposition matters because it separated three different questions:

• whether the stock-level relative-value logic was better in QQQ than SPY
• whether quote and loader issues were hiding the true branch quality
• whether the preferred option overlay should be the default single-leg structure or a higher-liquidity variant like qqq_single_volume15_v1

Once the branch was broken apart that way, the asymmetric signal became much easier to see.

Evidence / Results

The core evidence from the dispersion backtest now looks like this:

• strict-causal rerun: QQQ single +20.81% on 5 trades, vertical +10.85% on 8 trades
• patched combined rerun: single +5369.05 on 7 trades, vertical +8150.18 on 11 trades
• direct Stage B QQQ replay: qqq_single_base 9 trades, +44537.92
• direct Stage B QQQ replay: qqq_single_volume15_v1 10 trades, +32998.35
• positive-unlock replay: qqq_single_base 9 trades, +9079.91
• positive-unlock replay: qqq_single_volume15_v1 17 trades, +326.53

The symbol-level interpretation from those results is clear. QQQ carried more of the dispersion signal. SPY remained weak, sparse, or negative often enough that the combined view was less informative than the decomposed one.

What Worked

What worked was the symbol decomposition and the later dynamic-overlay repair. The repo learned that dispersion strength was not evenly distributed across sleeves, and it stopped pretending otherwise. It also learned that fixed overlays could create stock-to-option drift once the upstream winner changed. When overlays were generated dynamically from the selected stock seed, the branch became cleaner and the preferred winner changed from qqq_single_volume15_v1 to qqq_single_base.

This is exactly the kind of result a scientific process should produce. It identified a winner and clarified why the earlier combined branch was noisier than it should have been.

What Failed

What failed was symmetry. SPY did not cooperate in the same way QQQ did. That is an important negative result because it prevents a family-level victory claim that the evidence does not support.

Sample size also failed to clear the promotion bar. The repo was disciplined enough not to promote the QQQ sleeve despite large-looking totals like +44537.92 on 9 trades. That restraint matters. Large totals on thin samples are hypotheses, not admissions tickets.

Takeaway

The main lesson from this dispersion trading backtest is that the signal was not symmetric across QQQ and SPY. QQQ carried the economic value, SPY did not, and the combined view only made sense once the branch was decomposed and the overlay path was repaired.

If you want the strategy-logic side of this family, Relative Strength Breakout Strategy: Testing Proxy-Based Intraday Breakouts With QQQ and DIA explains the proxy-relative design. If you want the near-miss that still failed admission, Why c4 Was Parked: A Dispersion Strategy That Improved But Still Failed the Portfolio Gate is the next step. Join the research log to get the next backtest and failure report.