SYNAPSEWe publish the failures. That's where the rigor shows.
Every idea is specified before it runs, tested point in time with no look-ahead, and judged on sessions it never saw. Most don't survive. These are the dated verdicts, and the process a rule has to pass before it trades. The full write-up is a working paper: download the PDF.
We tested the indicators you were taught.
Each one below was run point in time with no look-ahead and judged on sessions it never saw; the recent studies were pre-registered in writing first. We publish the failures because that is where the rigor shows.
chart_patterns0 passEngulfings, stars, wedges, double tops: no pattern's out-of-sample range sits above a coin flip, even where a single rate looks high on a small sample. The textbook bearish ones fade more often than they follow through (evening star: 37.7%).
1,249 /ES sessions · 59,954 prints · 2021 to 2026rsi · macd · stoch · adxdeadThe RSI gate that looked brilliant was just picking puts (134 of 143 fires). And at a 0DTE entry, MACD(12,26,9) usually isn't defined yet: it was on only 104 of 360 entries.
point in time on 5-minute /ES · three regime splitscombine_everythingAUC 0.502Fifteen signals in a machine-learned model, walk-forward. In sample it looked like a printer. Out of sample its accuracy fell below simply being long.
857 out-of-sample sessionsauction_market_theory8 of 8 coinsValue areas, acceptance, rejection: eight independent constructions of AMT direction on /ES, every one a coin flip. Its range read survives; its direction read does not.
built, tested, closedthe_strat · ictcoinsThe Strat's 2-1-2 called direction 53.6% of the time, ICT fair value gaps 51.6%, liquidity sweeps 45.9%. Re-tested on real trade outcomes, as a filter and as an entry signal: dead both ways.
2024 to 2025 /ES · 360 tradesdirection_breakeven59% neededA 15-minute at-the-money 0DTE trade needs 59% direction accuracy to break even. No read we have measured gets there, ours included.
n = 1,212 sessionsWhat survives is structure and timing, not prediction. The one signal that passed, a chop filter on how efficiently price is moving, runs in the engine today. That is what Synapse is built on, and why it tells you when it is standing down.
Don't take our word for it.
We figured you'd be skeptical, so the desk runs your favorite systems live. Turn on the Strat, ICT, opening-range breakouts, VWAP, EMA clouds or auction market theory mid-session, and their entries land on the same chart as the engine's calls. Compare them yourself while the day plays out.
Recorded session, Thu Oct 1, 2026, replayed. Entries only, nothing scored. Not advice.
How we found the edge, and what it looks like in the backtest.
If you build models for a living, this is the part the rest of the site skips: the data, everything we tried, how the rules were chosen, and the hypothetical numbers that came out the other end, year by year, with the caveats next to them.
spxw_chains1,704 sessions of intraday SPXW option chains, Jun 2018 to Oct 1, 2026. About 100 strikes a snapshot, calls and puts, bid and ask, every five minutes.open_interest1,719 sessions of prior-day open interest, the input to the 09:35 dealer-pin read.es_tape/ES futures candles drive every directional read. SPX direction is never inferred from SPY.Tested and closed Survived and runs in the engine
The filters cost about 4% of the total and cut the worst drawdown by 35%. That trade, return for path, is the one we keep making. Pricing every fill at the ask instead of the mid takes another 15.6% off the total.
| Phoenix · butterfly | Peregrine · directional | |||||||
|---|---|---|---|---|---|---|---|---|
| Year | Trades | Win rate | Avg / trade | Profit factor | Trades | Win rate | Avg / trade | Profit factor |
| 2022 | 62 | 75.8% | $332 | 3.10 | 55 | 47.3% | $181 | 1.23 |
| 2023 | 169 | 77.5% | $18 | 1.09 | 60 | 55.0% | $409 | 1.80 |
| 2024 | 132 | 82.6% | $93 | 1.48 | 42 | 52.4% | $352 | 1.63 |
| 2025 | 110 | 83.6% | $205 | 1.80 | 39 | 64.1% | $2,155* | 6.92 |
| 2026 YTD | 56 | 82.1% | $140 | 1.47 | 6 | 66.7% | $2,097 | 5.51 |
| All | 529 | 80.3% | $126 | 1.58 | 202 | 54.5% | $722 | 2.28 |
*One session carries Peregrine's 2025: Apr 9, 2025 made $36,439 on one contract. Without it, 2025 averages $1,253 a trade at 63.2%, and the full window $545.
Phoenix wins often and small
Peregrine wins half and big
Now on trial: a language model. It reads the morning's news and gives a second opinion on each day's call, in shadow, with no say in the trade. It can only be judged going forward, because a model trained on market history may remember the answers to a backtest. Its test was written down before its first verdict, and the result gets published either way.
Try Signal free for 7 trading days.
$0 today. Cancel before the trial ends and you pay nothing.