Data basis: DAX and FTSE, 15 Apr 2015 – 05 Jun 2026, 2,851 and 2,824 usable days respectively (M1 data). Push = 08:59 close minus 08:00 open Berlin, normalised by the 20-day session range; strength as rolling terciles over 250 days, all from past data only. Part 1 is pure price structure, no costs. Part 2: all nine EU setups of our episode basis, 41,668 episodes, trailing stop BE 0.5 / TS 1.0 / step 0.5, net of spread and slippage. t-values over day clusters (DAX and FTSE on the same day are one draw). No trading recommendation.
The question came from live trading: when the EU indices have already run 50 or 80 points in one direction in the hour before the cash open — does follow-through come next, or does the market take the push back?
The follow-through camp says the push shows the direction of the day. The retracement camp says pre-cash trading is thin, and the cash open corrects it. We measured both — first as price structure, then as a filter on real setup trades. The answers differ, and that is the actual finding.
1. As a direction forecast: neither
All values signed in the push direction and in units of the session range (ATR). Positive means follow-through, negative means retracement.
| Push strength | Market days | Continuation 09:00→10:00 | 09:00→11:00 | 09:00→close |
|---|---|---|---|---|
| LOW (quiet) | 1,900 | −0.000 (t = −0.1) | −0.000 (t = −0.0) | −0.003 (t = −0.2) |
| MID | 1,875 | +0.010 (t = +1.3) | +0.011 (t = +1.1) | −0.004 (t = −0.2) |
| HIGH | 1,900 | +0.015 (t = +1.7) | +0.023 (t = +2.1) | +0.026 (t = +1.4) |
| Top decile | 601 | +0.024 (t = +1.3) | +0.027 (t = +1.2) | +0.044 (t = +1.2) |
A single value reaches |t| ≥ 2: the two-hour drift in the HIGH bucket. In DAX points that is 4 to 5 points — at the cost threshold. With around 30 t-values computed, one to three hits are the random expectation. The top-decile bucket, which would have to be stronger if the effect were real, is not. CAC confirms the two-hour value (t = 2.7 without clustering), SMI does not (t = 1.1).
The mechanics behind it, HIGH bucket:
| Metric | Value |
|---|---|
| Max continuation in push direction (median) | 0.449 ATR |
| Max counter-move (median) | 0.425 ATR |
| Days retracing > 50% of the push | 82.4% |
| Days erasing the push completely | 68.7% |
| Race: +50% extension before −50% retracement | 54.1% / 45.7% |
| Close in push direction | 51.5% (LOW bucket: 50.1%) |
Continuation and counter-move are almost symmetric. Two thirds of days undo the push completely — retracement is the rule. But the rule says nothing about the direction afterwards: 51.5% closing in the push direction is a coin flip.
2. As a setup filter: a different finding
Now the same daily feature applied to real trades. Each episode gets two attributes: the push strength of the day, and whether the trade goes with or against the push direction. The push is complete at 09:00; all nine EU setups trigger later.
| Cell | n | avgR | t | Δ baseline |
|---|---|---|---|---|
| Baseline, all EU setups | 41,668 | +0.122 | +17.0 | — |
| Push HIGH & WITH | 7,194 | +0.199 | +6.6 | +0.077 |
| Push MID & WITH | 6,988 | +0.160 | +5.7 | +0.038 |
| Push LOW (quiet) | 13,982 | +0.106 | +9.2 | −0.016 |
| Push MID & AGAINST | 6,810 | +0.080 | +2.2 | −0.042 |
| Push HIGH & AGAINST | 6,694 | +0.075 | +0.3 | −0.046 |
A monotonic gradient from WITH to AGAINST, spread 0.124 R. The against cell is statistically indistinguishable from zero — no man's land after costs. And the WITH filter lifts all nine setups above their own baseline, from +0.028 (the reference-candle setup, which already carries a filter of its own) to +0.206 (a morning fade setup). Averaged across setups: WITH +0.087, AGAINST −0.054.
How does that square with part 1? A day can close in the push direction or not — that is the coin flip. But a setup trade does not live until the close; it lives in the hours after its trigger with a trailing stop. Whether the market runs cleanly in one direction during those hours is a different question from where it stands in the evening.
3. Three counter-checks
Nine of nine setups positive is too clean to believe unchecked. Conditioning on daily features has died three times in this workspace on the same mechanism: "the setups align themselves — the trend side breaks first."
C1 — Is the push just the last candle? The last ten minutes of the window (08:50–09:00) are the reference candle of one of our setups. The two segments are independent at corr = −0.025 and 48.4% direction agreement, and each works on its own: 10-minute candle WITH Δ +0.097 (t = +8.6), early hour 08:00–08:50 WITH Δ +0.048 (t = +5.0). Holding the candle direction constant, the early hour still carries — and exactly where it counts: when the candle is AGAINST the trade, a strong early hour WITH the trade lifts the cell from +0.047 to +0.121 (n = 3,523, t = +2.3). When the candle is already aligned, it adds almost nothing (Δ +0.024). The candle is the main carrier, the hour before it the safety anchor.
C2 — Do the setups align themselves? Long share on upward pushes 55.9%, on downward pushes 52.8%, overall 54.6%. The HIGH bucket splits 51.8 / 48.2% into WITH and AGAINST. The setups do not preferentially trigger in the push direction — AGAINST is not a different trade population but the same setup catalogue on the other side. The filter works both ways: long Δ +0.096 (t = +6.6), short Δ +0.052 (t = +2.7).
C3 — Does it hold in both halves? 2015–2020: WITH Δ +0.086 (t = +4.6), AGAINST Δ −0.056. 2021–2026: WITH Δ +0.067 (t = +4.7), AGAINST Δ −0.036. Per setup and half, WITH-push beats its own baseline in 17 of 18 cells; the sole exception is one morning breakout setup in the second half.
4. What this means
Two measurements, two answers, no contradiction. As a forecast of the day's direction, the pre-open push is worthless — it belongs in the same row as our finding that the first candle does not predict the rest of the day. As a condition for a setup trade it is the opposite: the difference between trades with and against a strong push is 0.124 R, consistent across setups, directions and halves of the sample.
The reading is not "buy the push". It is: on days with a strong pre-open push, the counter-direction has no measurable expectancy left. That is an exclusion criterion — the same structure we find in the catalyst study and in volatility: daily features are good for filtering out, not for confirming.
Two thirds of the effect sit in the 10-minute candle, which is partly already in our setup catalogue. The new contribution of this study is the hour before it — and its value lies exactly in the cases where candle and trade disagree.
5. Limits
- In-sample across the full period. The half split is pseudo-OOS, not walk-forward; the effect was found on the same data it was checked on.
- Daily feature, coupled cells. The push hits all setups of a day at once; the cell n figures overstate the effective sample. That is why only the clustered t-values count.
- Multiple testing. 9 setups × 6 cells; at |t| ≥ 2 individual hits are chance. We therefore rely on consistency across setups, not on single cells.
- Part 1 without costs, pure price structure. The pre-cash hour is thin trading with wider spreads.
- Not tested: other push windows, push as a size regulator instead of a yes/no filter, US indices, transfer to CAC/SMI as tradeable markets (there only as a price-structure cross-check).
- CFD data, ending June 2026. The setup family is ours; absolute levels are optimistic, relative comparisons robust.