Twelve Rule-Session Cells on ES, Twelve Losses, and the One That Looked Like an Exception

Every intraday backtest is really a backtest of a session. We ran four rules against the Asia, London and New York sessions on twelve and a half years of S&P 500 futures — same code, same costs, only the hours change.

All twelve cells lose money.

An earlier version of this article reported one exception: a trend pullback paying +$46.90 a trade in New York at t = 13.74, decaying neatly to +$8.16 in London and +$0.28 in Asia, in the same order the sessions rank by volume. It was a tidy story and it was wrong. The number came from a bug in our own fill engine, and the corrected rule loses about $24 a trade in every session.

The correction is more interesting than the result it replaced, because it removes the explanation too.

The rules

ES 5-minute bars, 2 January 2014 to 10 September 2026. Each session is sliced first, then all four rules run on the slice with no cross-session carry.

• Opening range break. First 6 bars of the session set the range; the first break either way is the entry, filled at max(level, open) for longs and min(level, open) for shorts. Out at the session’s last close.
• Trend pullback. EMA-50 against EMA-200 on the close sets direction. Any bar closing with the trend marks its low (longs) or high (shorts) as a level, set once and never moved. A limit rests there for up to 24 bars. Stop 1× ATR(14), target 2× ATR, out after 48 bars.
• Mean reversion fade. Fade the first touch of 2 standard deviations from the session’s expanding mean. Stop 1.5 standard deviations beyond entry, target the mean.
• First-hour momentum. Direction of the first 12 bars, entry at the 13th bar’s open, out at the session’s last close.

One contract, $4.50 commission plus two ticks of slippage, $29.50 a round trip. A bar containing both stop and target is scored as the stop in every rule.

The grid

Session	Rule	Trades	Win rate	Per trade	Total	t
New York 09:30–16:00	Opening range break	3,209	47.8%	−$28.87	−$92,628	−1.13
New York 09:30–16:00	Trend pullback	21,136	34.9%	−$23.99	−$507,048	−8.31
New York 09:30–16:00	Mean reversion fade	2,913	48.5%	−$0.78	−$2,269	−0.09
New York 09:30–16:00	First-hour momentum	3,215	49.4%	−$11.90	−$38,255	−0.49
London 03:00–08:00	Opening range break	3,195	43.4%	−$42.80	−$136,740	−3.96
London 03:00–08:00	Trend pullback	14,545	35.9%	−$26.58	−$386,576	−12.59
London 03:00–08:00	Mean reversion fade	2,988	50.7%	−$5.76	−$17,205	−1.32
London 03:00–08:00	First-hour momentum	3,205	44.0%	−$40.42	−$129,560	−3.80
Asia 20:00–03:00	Opening range break	3,798	46.7%	−$29.52	−$112,104	−1.13
Asia 20:00–03:00	Trend pullback	19,243	37.4%	−$23.83	−$458,559	−16.99
Asia 20:00–03:00	Mean reversion fade	3,691	43.6%	−$22.55	−$83,222	−8.16
Asia 20:00–03:00	First-hour momentum	3,806	45.9%	−$9.79	−$37,264	−0.38

Twelve bars, all below the line. The closest to zero is New York’s mean-reversion fade at −$0.78 a trade, which on 2,913 trades and t = −0.09 is a coin landing on its edge rather than a finding. First-hour momentum in Asia (−$9.79, t = −0.38) and the opening-range break in New York (−$28.87, t = −1.13) are the other two that are not statistically distinguishable from zero, and none of the three is distinguishable from zero in the helpful direction either.

London is the worst session for three of the four rules, and by some distance: −$42.80 on the opening-range break and −$40.42 on first-hour momentum, both significant.

What the correction removed

The original article’s argument was that the trend pullback decayed with session volume. It had the numbers to support it — +$46.90 in New York, +$8.16 in London, +$0.28 in Asia — and the sessions rank by median five-minute volume as 11,674, 1,230 and 452 contracts. Three points in the right order is a persuasive picture, and the mechanism was plausible: a rule that rests a limit needs a counterparty at that limit, and in a 452-contract session there are fewer of them.

Here is the same rule with the fills resolved honestly:

Session	Median 5-min volume	Per trade
New York	11,674	−$23.99
London	1,230	−$26.58
Asia	452	−$23.83

A 26-fold range in volume, and a $2.75 range in outcome. The rule loses the same amount everywhere, and the thinnest session is not the worst one. Whatever is wrong with the trend pullback has nothing to do with how many contracts trade while it is running.

That is worth stating plainly because the liquidity story was ours and we believed it. It survived because the effect it explained was an artefact, and a good explanation attached to an artefact is harder to dislodge than a bad one — it makes the number feel earned.

The bug, briefly

Every test in this series that rests a limit order scanned each level’s own future to find its fill, then blocked new signals until that trade’s exit:

for i in signals:
    if i <= busy: continue
    for j in range(i + 1, i + 25):     # walk THIS level's future
        if touched(j): take(j); busy = exit_index; break

A level marked at bar 10 and filled at bar 40 sets busy past a level marked at bar 12 that would have filled at bar 15. The run keeps the later fill and discards the earlier one, which is a choice that needs the future. And it is not a random error: levels that take a long time to fill are the ones price travelled far to reach, so the scan systematically kept the deep pullbacks and deleted the shallow ones that get stopped.

The trade counts show it. New York went from 17,160 trades to 21,136; Asia from 15,856 to 19,243. The correction adds trades rather than removing them, and the added ones are the losers.

All fills now come from one shared engine that walks bars forward once, lets levels compete, and gives the book to whichever is touched first.

What this article is still for

The design is the control, and it survives intact. Running identical code and identical costs across three sessions is what isolates the session — there is no separate placebo here because the other two columns are the placebo.

What it was originally used to prove was that the pullback was a cash-session effect. What it proves now is narrower and less comfortable: the session is not the variable. Four rules, three sessions, twelve cells, and the spread between the best and worst session for a given rule never explains the sign. Everything loses in New York too.

We did not run random-entry placebos per cell here. The trend pullback article carries those, and they now lose as well.

Methodology: ES continuous front-month, 5-minute bars resampled from our own trade prints, 2 January 2014 to 10 September 2026. Sessions defined in ET on the bar’s own timestamp: Asia 20:00–03:00 (262,540 bars, 3,877 sessions), London 03:00–08:00 (192,251 bars, 3,239 sessions), New York 09:30–16:00 (247,335 bars, 3,643 sessions); each session sliced first, then all four rules run on the slice with no cross-session carry. Resting limits are resolved chronologically across all live levels, first touch takes the book, ties broken toward the level nearest the bar’s open; the superseded figures quoted here come from the same script before that engine replaced the per-level forward scan. A bar containing both stop and target is scored as the stop in every rule. One contract, $4.50 commission plus two ticks of slippage per round trip, $29.50 on ES. t-statistics are one-sample against zero on per-trade P&L. Median five-minute volume is taken across all bars in each session slice.