Research 9 min read

58,000 Simulated Trades, Every Level Family — and Nothing Beats Random Levels

After two studies found weak, in-sample edges around our levels, we did what a sceptical reader would do: threw every level family, every interaction and every conditioning variable at the data, and then asked the only question that matters — does the best thing we found beat the best thing a grid of random levels finds?

The short version

No. Across 460 unconditioned cells and 3,355 conditioned ones, the best real cell had a t-statistic of 2.68. The median best cell of a random-level grid had 2.80, and beat the real one in 14 draws out of 20. With conditioning, 3.79 against 3.81. On 84 sessions, the grid contains nothing distinguishable from chance — including the cells that look spectacular in isolation.

What "everything" meant

58,112 simulated trades on 78 RTH sessions of NQ 1-minute bars, with the levels the Indicator actually published each minute.

The cells that "stood out"

CellTradesWin ratePts / tradet-stat
Fade the upper VZ2 band approached from below, stop 40 / target 806459%+21.42.6
Same, stop 60 / target 1205060%+34.32.7
Same, transitional regime, stop 40 / target 604576%+29.03.8
Sweep and reclaim of the Battle Zone low, stop 20 / target 406157%+9.62.2

Any of these would make a convincing landing page. The first one even survives its own checks: positive in June, July and August, robust to the touch distance and to the stop size, confidence interval above zero by day. And it is exactly what the best cell of a random grid looks like.

The arbiter

Judging a cell by its own p-value is how backtests lie: search three thousand cells and a few will clear any threshold. The honest control is to rerun the entire grid on levels that cannot contain information — the levels of another day, shifted onto today's price — and record the best t-statistic each random grid reaches. Twenty grids, the same cells, the same conditioning. That distribution is the bar. A real cell only means something if it clears what the best random cell reaches.

Real gridRandom grids (median)Random grids that beat the real one
Best t, unconditioned cells2.682.8014 / 20
Best t, with conditioning3.793.8110 / 20
Cells with t ≥ 2.5149
Cells with t ≥ 352.5

One practical trap for anyone repeating this: permute the levels without re-centring them on the day's price and the random grids barely trade — prices differ by hundreds of points from one day to the next — which makes the null look weak and the real cells look strong. We caught that on the first run and fixed it before drawing any conclusion.

What survives, and what it means for the levels

Descriptively, the same asymmetries kept appearing across all three studies: continuation after acceptance beats fading at contact; rejections worked at the top of the day's range and not at the bottom; buying the Put Wall at contact lost 17 points per trade; buying an upside break of Focus 1 lost 14. Those are consistent, and they are what the Academy has always taught — read the regime, respect acceptance, never fade against the drift. What none of it is, on this sample, is a mechanical edge.

So the two rules we froze earlier are not "the edge". Their in-sample t-statistics sit inside what the best random cell reaches. They are hypotheses with every parameter written down, measured forward on the level statistics page, judged at 200 trades each. That is the only kind of evidence we will show, and the reason our levels are sold as context.

The levels, with the numbers next to them

The Indicator draws the map; the statistics page shows how it behaved; the Academy teaches the reading. Context, never signals.

Start the 7-day free trial — $9.99/mo

Method notes, for the people who will check

Disclosure: GEX Levels operates the Indicator and Education Library. This article is educational content about our own research, not financial advice, and past behaviour of any level is not a prediction of future behaviour.