Ran the first proper test of the volatility-squeeze breakout concept across all four Kraken pairs. Kraken’s REST OHLC caps at roughly 7.5 days of 15-minute bars, which is why every result on file was n=4–6 — a sample where the standard error is about half an R and nothing is distinguishable from anything. Pulled 52 days per pair from a second data vendor instead and re-ran.
Result: no edge. Pooled across three pairs, 109 trades, gross expectancy −0.05 R with a 95% CI spanning zero.
The number that settled it: a fixed 2R target breaks even at a 33.33% win rate. The three pairs came in at 33.3%, 30.0% and 32.6%. On BTC the gross expectancy was exactly zero — twelve winners at +2R against twenty-four losers at −1R. That is what a coin flip looks like once you draw an R-structure on it. The breakout was producing geometry, not direction.
Swept 45 parameter combinations. The best one failed its uncorrected 95% CI, before any multiple-comparisons adjustment was applied. So there is nothing here to tune toward, and tuning is off the table.
The bots stay running. They use public endpoints only — no keys, no signing, every fill logged as simulated — so accumulating real out-of-sample bars costs nothing, and out-of-sample data is the one thing that can move a result the in-sample sweep cannot. This remains the Kraken candidate; MarketCoach is the NQ strategy. The distinction being drawn is between running a candidate and believing it works, and only the second is ruled out today.
The bar is set now, in advance, so it cannot be moved later: a win rate clearing the 33.33% break-even line over 100+ forward trades, at a realistic cost basis.
That last clause turned out to matter more than expected. The backtest prices at spot taker rates, which cost 0.90 R per trade against a gross expectancy of −0.05 R — the strategy would have to win 63% of its trades just to break even. On the exchange’s futures maker schedule the same trades cost 0.04 R. A forward test graded against a cost model that forecloses its own conclusion is not a forward test, so the venue decision comes before any more accumulation.
Worth recording that the code was clean. Entries booked at a real traded price, not a synthetic level; costs correctly applied; the rolling-range contract honoured. An independent re-implementation driven by a different data vendor reproduced the published trade counts and win rates exactly. A null from a good process is a result, not a failure.
Two governance findings from the same pass:
Found and fixed a divergence between the MarketCoach research harness and the live strategy, then re-graded everything downstream of it.
The live strategy re-anchors its session state twice a day — at the futures session open and at calendar midnight — clearing the session VWAP, the daily signal count, the signal cooldown, the whipsaw window and the opportunity lifecycle together. The harness re-anchored once, on the calendar date only. For the nine hours between the session open and midnight the harness was carrying the previous day’s accumulated volume and stale counters while the live code had already reset. The VWAP side is a term in a hard five-way AND, so the divergence added and removed signals outright rather than merely shifting them, and the entire Asia session sat inside the affected window.
The correction costs more than expected. An earlier partial patch — VWAP only — looked mild. The faithful fix, which re-anchors every piece of state the live code resets, is materially worse: expectancy down by about a quarter, positive years down from 7/7 to 5/7, and the two years that were previously positive only by rounding both flip negative. One of those is the largest year in the sample.
The strategy still clears the break-even win rate for its own payoff structure by about ten percentage points, so it is still producing genuine directional information. What it no longer does is survive a multiple-comparisons correction against the search budget already spent on it — roughly twenty distinct configurations across stop buffers, bias filters, moving-average lengths and direction slices.
What changes as a result:
The fix ships behind a flag that defaults to matching live, and reproduces every previously published number byte-identically when switched off, so nothing in the history becomes unverifiable.
Check the win rate against the break-even rate for the payoff structure before anything else. A fixed-target strategy breaks even at 1/(1 + average winner). If the measured win rate sits on that line, the setup contributes no directional information and every downstream statistic is describing noise. One division. It would have ended the VolSqueeze investigation on the first page, and it is the check that most clearly separates the two projects logged above.
A boundary condition written twice will eventually be written two ways. The session anchor existed in two places in one file, neither wrong on its own terms, disagreeing for months. Where live code resets a block of state on one condition, the harness must derive every one of those resets from a single expression — and a test must assert the two implementations agree. This is the same defect family as the bot/backtest divergence found in the Kraken project three days earlier.
A configuration that passed walk-forward testing failed once the price data was corrected.
The original test used front-month NQ with contract roll gaps left in, up to 282 points. On a back-adjusted series the relaxed configuration turned slightly negative. The shipped configuration stayed positive, but its edge more than halved.
Every number got worse once the roll gaps were removed, and only one of the two stayed above zero. The relaxed setup was reverted and the original configuration restored in the source.
A roll gap is a price move that no one could ever have traded. A test that includes them is measuring the data, not the strategy.