Cross-Sectional Intraday Reversal at Multiple Horizons

A current, costed read on cross-sectional intraday reversal for a point-in-time top-500 US-equity universe, measured at 5-, 30-, and 60-minute horizons on ≈373 million Alpaca SIP minute bars. The effect is statistically real but breaks even below half a basis point of round-trip cost — and the study doubles as a reusable cross-sectional evaluation template.

We replicate and re-measure cross-sectional intraday reversal — the tendency of recent intraday losers to out-perform recent intraday winners over the following minutes — for a point-in-time top-500 US-equity universe on 1-minute consolidated (SIP) bars over 2018–2025. The pipeline ingests ≈373 million minute bars across 529 symbols, sorts each h-aligned cross-section into deciles on the prior-h return, and measures the next-h equal-weighted decile spread for h ∈ {5, 30, 60} minutes.

The 30-minute long–short returns 11.3% annualized gross over the eight-year window. Pooling all 2,007 trading days, that edge is statistically distinguishable from zero (95% block-bootstrap interval [+1.2%, +20.7%], p ≈ 0.04) but economically negligible: it is worth only ≈0.45 bps per sort against ≈1.7× round-trip turnover, so it breaks even at only ≈0.26 bps of round-trip cost and the tradeable signal decays inside a single sort interval (no resolvable exponential half-life). At the 5-minute horizon the effect is enormous (≈148% annualized, CI [+126%, +169%]) but is essentially a bid–ask-bounce artifact concentrated at the session edges.

The headline universe is a point-in-time (PIT) monthly top-500 membership table — a continuous monthly snapshot derived from an S&P 500 historical-components source spanning 97 month-ends from 2017-12-31 through 2025-12-31. Each snapshot holds exactly 500 names with complete coverage; membership for any trading day is resolved as-of the most recent prior month-end, so no constituent change is used before it is known.

One-minute OHLCV bars come from the Alpaca SIP feed (full consolidated tape, history to 2018) for the union of all PIT members. Each symbol-day is reindexed onto the NYSE regular-session minute grid (timezone-converted to America/New_York, holidays and early closes respected); price-like columns are forward-filled at most one minute within a session, and any symbol-day with more than 5% pre-fill missingness is dropped. Per-minute log returns are computed within-session only, never crossing the overnight boundary. An earlier attempt with the Polygon free tier failed because that plan only covers ~2 recent years — a provenance lesson that drove the switch to Alpaca SIP for the full download.

Sign convention. We fix the primary reported metric to the reversal payoff: the next-h return of the prior-h losers (decile D1) minus that of the prior winners (D10). A positive value means recent losers subsequently out-perform recent winners.

Decile sort. For each horizon and session, sort timestamps step by h from the open plus a 15-minute open/close exclusion, requiring complete prior- and forward-h windows. At each timestamp the eligible cross-section (valid prior- and forward-returns, ≥10 names) is ranked on its prior-h log return and split into deciles with a deterministic quantile estimator; each decile’s equal-weighted forward return is measured.

Turnover and annualization. A turnover proxy tracks top/bottom-decile basket changes across consecutive sorts; we report round-trip turnover. Intraday long–short returns sum to a daily P&L and annualize at 252 days. Because each horizon fires many times per day (≈70 sorts/day at 5 min vs. ≈4 at 60 min), the 5-minute series compounds far more shots — a fact that inflates its annualized figure.

Inference, decay, and costs. Per-year and pooled full-sample 95% intervals use a 5-day block bootstrap (250 iterations, seed 1729). Decay is traced over a grid of sort-to-hold gaps g ∈ {0, 5, 10, 15, 30, 60, 120} minutes; the requested exponential half-life is reported but, as below, is not identifiable, so we lead with the non-parametric retained fraction ρ(g) = |R^rev(g)| / |R^rev(0)|. Each round-trip cost is charged against turnover and we solve for the break-even cost at which annualized net crosses zero.

At the 5-minute horizon the gross reversal is enormous and remarkably stable — a mean of ≈148% annualized with a Sharpe of 5–10 every single year. At 30 minutes the mean drops to ≈11% with Sharpe below 1, and at 60 minutes to ≈6% with Sharpe ≈0.5 and several outright negative years. The monotone collapse of the effect from 5 to 60 minutes is the central empirical pattern, and it is the first hint that what survives at short horizons is microstructure rather than information. (The headline 11.3% is the simple average of the eight yearly figures; the day-weighted pooled mean is 11.28%.)

The large 5-minute Sharpe ratios are overwhelmingly significant. The economically interesting 30-minute book is marginal year-by-year — its 95% bootstrap interval straddles zero in 7 of 8 years, with 2023 the lone exception — but that conflates a genuine null with small per-year samples. Pooling all 2,007 trading days resolves the question: the full-sample 30-minute effect excludes zero ([+1.2%, +20.7%], p ≈ 0.04), the 5-minute effect is hugely significant, and the 60-minute effect is not distinguishable from zero ([−2.7%, +13.9%], p ≈ 0.10). The honest statement: the 5- and 30-minute effects are real, the 60-minute effect is noise, and "real" at 30 minutes still means an annualized edge whose lower bound is barely positive.

Is the spread a smooth cross-sectional reversal or an artifact of the two extreme deciles? The hero figure plots the equal-weighted mean next-h return for each prior-return decile, averaged over every sort in 2018–2025. The 5-minute profile is cleanly monotone: subsequent return falls steadily from D1 (+0.44 bps, biggest prior losers) to D10 (−0.40 bps, biggest prior winners), the textbook reversal staircase. At 30 and 60 minutes the reversal survives mainly at the extremes (D1 highest, D10 lowest) while the middle deciles are noisy — increasingly so at 60 minutes, mirroring that horizon’s statistical insignificance. The implied D1−D10 spread reconciles exactly to the per-sort gross edge below, an independent check on the long–short construction.

How fast does the edge decay as the holding window is delayed by a gap g? At the 5-minute horizon only ≈1.6% of the gap-0 edge survives a single 5-minute delay; at 30 minutes the retained fraction bounces between 6% and 56% across gaps — the trace of noise, not of a smooth relaxation. The signal is a gap-0 spike that falls to the noise floor immediately and stays there, non-monotonically.

This is where the effect dies. The book turns over ≈1.7× per sort and fires up to ≈70 times a day, so a per-sort gross of a fraction of a basis point cannot survive any realistic spread. Even the smallest tested cost (2 bps round-trip) drags the annualized 30-minute net deeply negative; the mean break-even cost is 0.49 bps (5 min), 0.26 bps (30 min), and 0.35 bps (60 min) — all far below realistic execution costs and below the bar-based spread proxy itself. The loss is dominated by turnover × cost, not by the small, noisy gross edge.

Survivorship. Running the identical analysis on a naive "current top 500" membership changes the 30-minute mean only marginally (11.31% PIT vs. 11.50% naive). The expected teaching artifact — a large PIT-vs-naive gap — does not appear, because the PIT source is an S&P 500 component proxy that overlaps heavily with the current list and because intraday reversal is a within-cross-section microstructure effect insensitive to slow membership churn. The lesson stands but is inverted: survivorship bias is small for this effect on this universe proxy; it would re-emerge for slower, return-level factors or a true market-cap universe with more turnover.

Open/close microstructure. Including all minutes lifts the mean 5-minute annualized return to ≈217% (vs. 148% with the 15-minute exclusion) while the 30-minute mean barely moves (14.0% vs. 11.3%) — the raw 5-minute effect is materially an open/close phenomenon. Spread filter. Dropping the widest-range names cuts the 5-minute gross magnitude by 20–30% and leaves 30/60-minute numbers broadly similar, consistent with the short-horizon effect being concentrated in wider-spread names where bid–ask bounce is largest.

The results paint a coherent picture. Intraday reversal in post-2018 US large caps is statistically present and correctly signed, but its tradeable content is tiny and horizon-dependent in exactly the way a liquidity-provision / bid–ask-bounce effect should be:

• Largest at the shortest horizon and the session edges. The 5-minute, gap-0 effect is huge in annualized terms but only ≈0.84 bps of signed mean per sort, is amplified by open/close minutes, and is concentrated in wide-spread names — the signature of quoted-spread mean-reversion rather than information-driven reversal.
• Evaporates with a one-period gap. Skipping even 5 minutes between sorting and holding collapses the 5-minute signal to ≈1.6% of its gap-0 magnitude — what you expect if the "edge" is the bounce off the price you would have transacted at.
• Cannot survive costs. Break-even is below 0.5 bps at every horizon, well under realistic spreads and fees, so the net strategy is deeply negative regardless of year.
• The annualized 5-minute figures are a compounding illusion. The 5-minute book fires ≈17× more often than the 60-minute book, so a sub-basis-point per-shot edge annualizes to triple digits. This is arithmetic, not alpha — and precisely the quantity that costs erase.

The honest headline is the negative one: a well-known effect, measured cleanly on current data, is confirmed statistically real but decayed below tradeability. That is a valuable baseline — it establishes what "not worth deploying" looks like and gives any future cross-sectional candidate a costed yardstick to beat. Just as valuable as the number is the reproducible, schema-validated, point-in-time, bootstrap-and-cost-aware template that any subsequent study can reuse and that any future alpha must out-earn on a costed basis.