Closing auction price discovery

I kept seeing the same late-day picture: SPX appeared to weaken around 15:50 ET, sometimes recovering near 15:55. The practical idea was straightforward: enter before the move and exit after it.

The research question had to be stricter: does the 15:50–15:55 window have a persistent bearish bias, and is any information available early enough to act on it?

The initial trade idea motivated the study. It was not treated as evidence.

An older 249-day sample showed a 55.02% bearish rate from 15:50 to 15:55, with a one-sided p-value of 0.064. It was suggestive, but it did not meet the frozen evidence standard.

In the newer 498-day primary sample, the bearish rate was 47.99%, the one-sided p-value was 0.827, and mean return was +0.675 bps.

The periods were kept separate rather than pooled.

Result: no convincing unconditional 15:50 downside effect.

That rejection changed the project. The next question was no longer “how do I trade the pattern?” It was “why does this time of day look structurally different at all?”

Simple conditions did not rescue the original hypothesis. A predefined late-day pressure setup moved in the opposite direction. A descriptive Thursday effect failed on its validation block. SPY’s own Arca closing imbalance did not separate later SPX outcomes.

Rather than keep tuning thresholds around a weak pattern, I moved closer to the market structure of the close.

Is there a real closing-auction price-discovery event around 15:50, even if it is not an unconditional bearish anomaly?

Closing pressure is distributed across hundreds of constituents and multiple listing venues. A single SPY imbalance feed was not enough to represent that cross-section.

I built a point-in-time SPY constituent universe from historical filings, linked holdings to stable security identifiers and primary listing venues, preserved historical ticker and venue transitions, and combined NYSE and Nasdaq closing-auction messages into a capitalization-weighted Constituent Imbalance Signal, or CSI.

The signal used only eligible messages received between 15:50:00 and 15:50:11 ET. Outcome data remained separated until the signal protocol and coverage checks were frozen.

• 63,396,782 imbalance records acquired after historical mapping patches
• 464 of 464 normal dates passed the frozen 95% coverage gate
• 99.6438% minimum absolute SPY-weight coverage

The data system mattered because the research question was now about what information existed, when it existed, and whether a trader could have known it yet.

Once the full constituent signal was available, it no longer offered convincing predictive power for subsequent SPX moves. High-frequency ES data, however, showed where the visible 15:50 effect was coming from.

From 15:50:00 to 15:50:12, Sell-state ES averaged −2.50 bps while Buy-state ES averaged +2.28 bps. The Sell-minus-Buy separation was −4.79 bps.

From the frozen actionable time, T0 = 15:50:12, to 15:51, the additional separation was only −1.04 bps.

This is a timing result, not a profit claim. It says that the market was already absorbing the closing-auction information while the full constituent signal was still arriving.

A real mechanism can be visible in the data and still arrive too quickly to become a usable signal.

After T0, the remaining relationship was small but directionally aligned. Across 459 primary ES dates, Sell-state mean return from 15:50:12 to 15:51 was −0.537 bps and Buy-state mean return was +0.500 bps.

That is a statistical relationship. Execution is a different test.

Using one ES contract, the observed bid/ask, the signal direction, and a frozen $5 round-trip fee, the residual price relationship was not enough to support a fee-positive strategy.

Statistically detectable does not imply economically executable.

If most of the response had already happened by T0, the natural next question was whether earlier market behavior anticipated the closing imbalance.

It did, to a point.

Using the same frozen pre-close ES price-path feature, its association with the eventual Combined CSI rose as 15:50 approached: 0.112 → 0.224 → 0.243 → 0.279 from 15:40 to 15:49.

But at 15:49, the same feature’s Spearman correlation with the subsequent 15:50–15:51 ES return was only 0.013.

The market could anticipate part of the imbalance without providing a reliable later price direction.

Twenty-five registered pre-15:50 strategies and models were evaluated. None passed the full promotion gates.

If part of the closing imbalance was already anticipated before 15:50, the relevant quantity might not be Actual CSI. It might be the residual.

The leakage-safe construction is:

CSI Surprise = Actual CSI − OOF Expected CSI

Expected CSI was built out of fold with chronological training, training-only preprocessing, and model selection based only on CSI forecast error.

The decomposition clarified when price discovery occurred. It did not create an executable result.

The project allowed exploration, but it did not allow attractive historical results to bypass the research protocol.

Rules were registered before evaluation. Dates were never randomly shuffled. Model selection stayed inside training windows. Executable tests crossed the observed bid/ask spread. Cost sensitivity, minimum sample sizes, outlier checks, calendar stability, and multiple-testing controls remained in place even when a candidate looked promising.

Three separate registries ended in the same decision.

0 strategies advanced to forward validation

The project closed when the remaining historical evidence no longer justified another round of threshold, timing, or feature search.