eric@builds:~$ cat edges/scoreboard.md
Edges: what the machine found
An always-on search over millions of intraday futures rules (NQ, ES, GC, CL, RTY), judged the hard way: discovery years first, then one locked holdout year it may test once per idea family per week, a coin-flip control that must lose where the rule wins, and finally the broker's own fills. The scoreboard below is written by the machine itself. It includes the honest number of confirmed edges, which may be zero.
loading the feed…
The other way round: few hypotheses, declared first
The brute-force search above can only promote miracles: a 60-minute NQ move has a standard deviation near $1,800 a contract, so its p ≤ 0.0005 bar needs a Sharpe-4 edge. Real edges run Sharpe 0.5 to 2. So this table tests a short list of scheduled, causal effects with a published record (the overnight equity premium, turn of the month, pre-FOMC drift, post-announcement drift, intraday momentum, the gold fix…), one position per day, Bonferroni over the declared list only, discovery years then the locked holdout. Rows marked control are the matched non-event days.
| contract | hypothesis | days | $ / day | Sharpe | holdout | verdict | |||
|---|---|---|---|---|---|---|---|---|---|
| not run yet | |||||||||
"HOLDS" means worth a paper lane, never proven. Overnight rows hold 18:00 to 09:30 ET, which most futures prop firms allow (flat by 16:59 ET is never crossed). Dollars per contract after commission and two ticks of slippage.
Every rule that earned a holdout test, newest first. Open one for the evidence. Dollars are per trade on one contract after commission and two ticks of slippage, in a simulator — not real trading.
Strongest signals on the discovery years that have not yet spent a holdout
| rule | contract | horizon | n | $ / trade |
|---|
A high t-statistic here is cheap: it is the best of thousands. The holdout year is what counts.
NinjaTrader fill check
The same rules replayed inside NinjaTrader's Strategy Analyzer with its own fill engine. Appears once the first batch has run.
Signal first, exit second. Most "edges" found by brute force are the exit talking: a wide stop and a tiny target win 95 % of the time and still lose money. So an entry idea is first judged with no stop and no target at all, at fixed horizons of 15, 30, 60 and 120 minutes and to the close. Only an entry that makes money on its own earns a bracket.
- Discovery (all history before the holdout year): at least 60 trades, at least +$20 per trade, p ≤ 0.0005, positive in 4 of 5 time blocks, and at least +$20 better than the same entries with a coin-flipped direction.
- Bracket pick: the stop and target are chosen in ATR units as the best 3×3 neighbourhood on the grid, a plateau rather than a spike, and the chosen bracket must pass stage 1 again on its own.
- Holdout: one test per idea family per week on the locked year, with a Bonferroni-corrected threshold. The rule is scored twice, with its bracket and with a time-only exit, so a bracket that flatters it on the discovery years is caught.
- Broker fills: survivors are replayed inside NinjaTrader with its fill engine, then paper-traded forward on a simulator before any real money is discussed. Nothing on this page places a trade.