I'm a pediatric surgeon. I've been trading ES for about three years, funded prop accounts mostly, and I've spent most of that time building systems that didn't work. Footprint imbalance scalper, a value area fade, a derivative-based wave detector, an ICT period I'd rather not talk about, a London session scalp that looked fantastic on bar data and died the day I got tick data. I've passed evaluations at about the published base rate and lost funded accounts at roughly the same rate as everybody else. So no, this isn't a "my method" thread.

What I do have is a habit from the other job. In medicine you write down the claim first, you fix the protocol before you look at the data, and then you publish whatever comes out, including the nulls. I started doing that with trading concepts instead of just trying them live, and I'm going to log the results here. Mostly because writing it for strangers forces me to be honest about n and about what I'm actually measuring.

First one. The value area fade. You know the trade.... price pokes above yesterday's VAH, fails, you short it back to the POC. Limit at the boundary on the failed probe, stop just past the probe extreme, first target +2, runner to POC, stop to breakeven after the first target. That's the version as it's actually traded, I took it from a written spec, not from my own head.

Data: 66 RTH sessions of 2026 ES, tick level. Then a file I hadn't opened before, 80 sessions of 2022, May to September, bear market, 77 million prints. Rules frozen on the first file before I touched the second.

The rule as written loses in both years. 2026: the filters were so tight it only fired 5 times in 66 days, gross -$212, and the POC target was hit zero times. 2022: 36 trades, 47% winners, gross -$1,988, POC hit 5 of 36.

The reason is boring and structural. The breakeven move puts the stop at entry +/- a tick, and one 1-minute ES bar has a median range of 2.5 points, so the runner is dead before it starts. And the POC target is about 19 points away on average against a 3 point stop, which is a 6R trade that pays off a few times a year. 80% of the trades ended between -1R and 0.

Then I did the thing I should have done first. Took every filter off and compared each failed probe at a value boundary against a random entry in the same session, same side, same holding time. In 2022 the probe beat random by about 2 points at 60 minutes (t = 6.2, n = 7,160) and about 4 points at 120 minutes (t = 9.1). 2026 showed the same direction on a smaller sample.

So the concept is real. A failed auction at the value edge carries information. The trade everybody wraps around it throws the information away. That was the lesson for me: the concept and the trade rule are two different claims, and I'd been testing them as one for three years.

What would change my mind: the probe-vs-random gap disappearing on a third, untouched year. Next on this one is a survival table of fixed targets from 4 to 12 points on the 2022 file, since on 2026 the expectancy crossed zero somewhere around 6-7 points but never got anywhere near significance.

I'll post the next one when it's done, probably the trend-day-vs-chop-day question (can you tell by 10:30? spoiler: on 243 sessions, no). If anyone wants a concept tested, state it the way it's actually traded, entry, stop, target, and I'll run it and post whatever comes out. I keep the full write-ups with every number at qumsiehtrading.com but the forum post will have the substance.

Not advice, obviously. I'm a surgeon who loses money carefully.

Second one. The trend day call. Everybody does some version of it: read the first hour, decide if it's a trend day or a balance day, trade the afternoon on that. I did it for a year and a half. So I tested whether it's possible at all.

243 RTH sessions, one year of 1-min ES, holidays and half days out. I measured the morning from 9:30 to a cutoff (tried 10:00, 10:30 and 11:00) three ways, how efficiently price travelled, a variance ratio, and how many breaks of the prior 20-bar extreme failed. Then asked if any of that predicts the same measures from the cutoff to 14:00.

Nothing. At 10:30, morning efficiency against afternoon efficiency correlates at +0.005. Variance ratio +0.08. False-break rate +0.03. All p above 0.2, and whatever small number showed up in the first half of the year vanished in the second half. Splitting mornings into chop/middle/trend terciles gives afternoon efficiency of 0.073, 0.071, 0.076. Afternoon false-break rate 78%, 79%, 80%. The afternoon doesn't know what the morning did.

Direction didn't carry either, 54.7% / 54.3% / 50.6% depending on the cutoff, and unstable.

The thing that convinced me the test wasn't broken: morning range predicts afternoon range at +0.57, p under 0.0001, stable in both halves. The pipeline finds a real relationship when there is one. Size carries over, character doesn't.

What ES actually is between 10 and 2 on basically every day: efficiency around 0.07, and about four of every five breakouts fail. That's the instrument. "Trend day" is a label you can apply at 4pm.

This one cost me a lot. It was two of the three steps in how I traded.... decide direction, decide regime, then trade the level. Only the third step survived testing. The practical consequence for me was to stop calling the day and just trade the level with the stop at a level, which is boring, and is the only part that works.

What would change my mind: any morning feature that predicts afternoon efficiency out of sample at p < 0.05 with both halves agreeing. I'd like to be wrong about this one.

Next: fair value gaps, on tick data, with the definition written down before I open the file.

Third one. Fair value gaps. This is the one I wanted to get to, because I spent a stretch of my ICT period trading exactly this and never measured it.

The trade as people actually take it: 3-candle gap, buy the retrace into a bullish gap (sell into a bearish one), stop past the gap, target 2R. I wrote the definition before opening the files: 5-min ES, RTH only, candle 3's low above candle 1's high, gap at least 1 point, limit at the near edge on the first return, stop a tick past the far edge, flat at the close.

Three windows, 82 sessions, 647 trades. Sep 17 to Oct 23 2026 was the main file. Jul 27 to Sep 16 and Mar 23 to Apr 17 I didn't open until the rules were frozen.

As traded: -0.55, -0.61, -1.10 points per trade after costs. Win rate 36%, 34%, 28%, against the roughly 33% a 2R trade needs before costs. About 75% of gaps get revisited the same session, so the "gaps get filled" part is real. It's what happens after the fill that doesn't pay.

Then the same check as Trial #1: every retrace into a gap against random entries in the same session, same side, same holding time. At 15, 30 and 60 minutes the gap entries were between -1.4 and +0.9 points vs random, 7 of 9 below zero, nothing significant. On 1-min bars it went the wrong way: gap entries did worse than random, t around -2.8 at 60 minutes. The gap gets run through more than it gets defended.

So, a null. Price fills gaps, but a gap isn't a level price reacts from any more than a random spot is. Unlike the value area probe in Trial #1, there was no information hiding under a bad trade rule.

Caveat, and it's a big one: no higher-timeframe filter. Nobody serious trades every gap. I tested the building block, not the full model. If you trade FVGs with a specific filter, write it the way you take it (entry, stop, target, filter) and I'll run it on the same three windows.

Next: killzones.

ES levels today: 7670 / 7651-57 / 7640 / 7626 / 7613-17

Fourth one. Stacked imbalances on the footprint. This is the one that hurt, because for a day I really thought I had finally found something.

The claim as people use it: two or more stacked buy or sell imbalances show aggressive size, and price respects that zone when it comes back.

What I ran: ES tick data, RTH 10:00 to 16:00, 15-minute footprint bars. Diagonal imbalance 2:1 with at least 120 contracts on the level, two levels stacked in the direction of the bar. Limit order at the near edge of the stack on the retest, 10-bar expiry, 8 point target, 5 point stop, one position at a time.

I built it on ESU26, 68 sessions. Swept about 6,500 parameter combinations on the first two months and kept August sealed until the end. On that contract: 150 trades, 58.7% winners, profit factor 2.2. The sealed month held. The neighbouring settings all held too, no knife edge anywhere. I was ready to trade it.

Then I ran it once, nothing changed, on contracts it had never seen.

ESH26 (Dec 2025 to Mar 2026, 67 sessions): 181 trades, 34.8% winners against 39.2% needed to break even, profit factor 0.83. Below break-even in every month, long side and short side.

ESU22 (Aug 2022): 49 trades, 32.6% against 38.5% needed, profit factor 0.66.

What was going on underneath: the 120 contract floor was doing all the work. Take it out and the win rate on the build contract drops from 58.7% to 40%. And per-level volume went from a median of 125 contracts in June to 728 in August inside that same sample, so the floor that made it work in June was barely filtering anything by August. It was fitted to one quarter. Same family of rules failed on NQ, the euro and gold as well, around 80,000 combinations in total before I stopped.

Caveats: neither fresh contract is significantly negative on its own, the edge just didn't show up. The backtest counted a touch of the limit as a fill, which is generous for a resting order sitting where size is waiting. And 2022 was a different volatility regime. None of that saves it, the calm 2025-26 contract failed too.

What would change my mind: the exact same rule, no retuning, beating break-even on two more fresh ES contracts.

Next one is killzones. Full write-up with all the numbers is on qumsiehtrading.com.