Start from a real result
Each one fills the form with a strategy the shelf already tested, settings and all. Change one thing, or send it as it is and read the report. A pass, a close call and two fails, on purpose: a fail teaches as much as a pass.
The morning move, NQPassed 90If NQ moved 0.4% or more from 8:00 to 9:30, trade against it; out at 10:15.716 trades · +$113,468 · 46 months flat at worstStart from this →
The overnight gap, RTYClose 52A gap of 0.3% or more from yesterday’s 4:00 close, traded back toward it; out at 11:00.2,178 trades · +$25,797 · 68 months flat at worstStart from this →
Opening range breakout, NQFailed 23The first close outside the first 15 minutes’ range, by 11:30; out at 3:30. It made money and still failed: the coin flips did as well.4,762 trades · +$20,005 · 1,603 of 5,000 coin flips as goodStart from this →
RSI 30 cross, ESFailed 11RSI(14) crossing down through 30 buys the bounce, 10:00 to 3:00; out an hour later.4,128 trades · -$175,738Start from this →
Or build it yourself:
Tap your setup
Last choices
Filters are how most strategies get fooled. Add a condition, the backtest looks better, add another, better still, and by the third one you've built something that only works on the past. So nothing here stops you, and the report keeps score: every condition is named, its settings count against your score, it counts as one more version of the strategy, and you see the trade count before and after each one.
Each filter is applied to the signal list at the signal bar or on completed prior days (no look-ahead). Each adds its tunables to the parsimony count, counts as one more try, is perturbed ±10% in Exam VI, and appears in a condition-by-condition table; the base-vs-filtered comparison is printed for the whole stack. Under 100 trades there is still no verdict.
A filter removes days. Fewer days fit easier and prove less. Its settings count against your score, and the report shows what it removed.
The range is the setup’s own, from the Strategy step. Only bars inside the range are read.
A list of days kept on the engine with its source and the date it was checked, so nothing in it is typed from memory. The report prints the list's name, how many days it holds, where it came from and when it was verified. Four lists today: the FOMC decision days, 2008 to now, scheduled and unscheduled, from federalreserve.gov; and the 8:30 a.m. ET release days from bls.gov, 2008 to now, as one list of both payrolls and CPI or as either on its own.
Provenance-gated CSV (date, source_url, verified_on on every row) resolved by name on the engine; embedded into the standalone at generation. Counts as one tunable.
"Fade" means bet the move reverses: it went up, you sell. "Follow" means bet it continues: it went up, you buy. Same trigger, opposite trade. The coin-flip test is what tells you whether either one is really better than chance.
Mode maps sign of move to side. Fade: down-move → long, up-move → short. Follow: the reverse. The coin flips are run on the same signals with random sides, so the side is the exact decision being tested.
Both sides, longs only, or shorts only. Picking one side roughly halves the trades, so the report will say if the sample got too thin.Side filter applied after signal generation; the day still counts as traded even if its signal was on the excluded side. G6 refuses a verdict under 100 trades.
The chart your strategy reads. Pick 3 and the averages, crosses and levels are computed on 3-minute bars, the way they'd look on a 3-minute chart. Fills, exits and costs still happen on the one-minute bars underneath, so nothing gets rounder or cheaper. Above 1 minute it counts as one more setting.One-minute bars aggregated by clock bucket (start-minute // N, per calendar day; OHLC first/max/min/last, V summed) for the signal pass only. Each signal fills at the first one-minute bar after the N-minute close; clock and hold exits are re-attached on the one-minute series; the coin flips, filters and cost stress see one-minute fills. Adds one tunable when N > 1.
One means the first signal of the day is the trade for that day, the way every strategy has been scored so far. More than one means the strategy can take a new signal after the previous trade is out on its clock, up to this many a day. It counts as one more setting. Strategies that get out at a clock time (opening range, time of day) can't re-enter, and the form will say so. A strategy you build from parts can, once it has a stop, a target or an exit on a signal: each next trade starts at the first signal after the last one is out.Re-entry cap. The family emits every qualifying signal in its window; a selection pass takes at most N a day and admits a new entry only after the previous trade's clock exit, which does not depend on the side, so the coin flips keep the same entries. Refused for clock-exit families A, C, E, P and for B. Adds one tunable above 1.
What your broker charges to get into and out of one contract, fees included. The shelf and every run without a figure pay $4.20. Yours is a defined cost, so the test can use it: a strategy that pays $2.00 a round turn keeps more of each trade than one that pays $4.20, and the report says so, printing the net at your figure and at $4.20. Slippage is different: nobody controls it, so it stays one tick a side on every run. Enter it on your first run; a run again at another figure counts as one more version, like any other change.COMMISSION_RT for this run only, 0.50 to 20.00 a round turn per contract, each exit leg paying it as before; the record carries the figure and the same trades at the standard. Slippage stays 1 tick a side; the 2- and 3-tick lines, the breakeven in ticks and where-costs-bite are computed at your figure. The coin flips pay what the rule pays, so their count does not move with it; the net, the year-by-year, the bootstrap range and the breakeven do, and the score can.
Per contract, in and out, fees included. Leave 4.20 for the standard every shelf result uses. A run again at another figure counts as one more version, so enter yours the first time.
The most important question on the form. A strategy is made by someone looking at charts, and the years they looked at can’t prove anything; the years before can. So pick the year the strategy first existed: when you started using it, or, if it came from a course, a seller, a book or a video, when they first taught or sold it. Everything earlier is scored as the honest test. Say 2024, and 2008 to 2023 is the part that counts. Not sure? Pick “2023, or not sure”, the line the shelf uses. A strategy known before 2010 has no years before it to test on here, and the report says so.
Declared contamination boundary: the earliest year the strategy’s author could have fitted it to data (your first use, or when a taught or sold strategy was first published). Years ≥ this are kept apart; the years before it are the ones the coin flips, the bootstrap range and the year-by-year check are run on (calendar years as windows, minimum 8). “Not sure” is 2023, the shelf’s boundary. Declared, not sworn; the report’s scope line says so. 2009 or earlier leaves no years before it to test on, which the report states plainly.
Every year before this one is the test. Years from this one on were already looked at, by you or by whoever made the strategy, so they don't count in its favor.
If you try ten versions of an idea and keep the best, the best one looks better than it really is, the same way the tallest of ten strangers is probably taller than average. Telling us how many you tried lets the report correct for it. Lying only makes your own report wrong. We also count submissions from your email, so the number used is the larger of the two.
Declared tries. The reported p is also shown counting every try, as 1 − (1 − p)^n, where n is the larger of the tries you declared and the submissions from this address plus one. The code file comes with a pass that also clears four more checks; the report lists the five. We can't recover your private search history; we can refuse to pretend it was zero.
About you, optional
Kept: your email, your level, your platform and the score with its headline numbers. The strategy goes the moment the report is sent.
Your strategy, written back to you
Read it once. This is exactly what gets tested.
Once you confirm, this strategy is locked. It runs exactly as written above. Any change afterwards is a new submission, counted as one more version.
Fixed for every run: one-minute fills, exits and costs (signals on the bar size you chose) · every day on file from 2008 (the data page names the last day) · $4.20 a round turn, or the commission you entered (the report then prints the net at both) + 1 tick per side slippage, fixed · market orders at next bar's open · unadjusted continuous data. All printed on the report.
What this test cannot see. Your read of the tape; which of these days you would have skipped; the fills you would really get. It tests the written strategy, nothing more. Your part is measured another way: the trade audit holds your own fills against this same strategy and shows, in dollars, what your choices added or cost. Audit your trades →