Crypto RSI(2) Dip-Buy With a Resting Limit Order — 1h

Status: Regime entry: worked on all three majors in 2018–22, on BNB only since · Regime #7 · Listed 2026‑10‑14 Instruments: BNBUSDT perpetual; BTCUSDT and ETHUSDT in 2018–22 only · Timeframe: 1h · Direction: long only Frequency: ~100 trades/yr · Fees: 0.02% maker in, 0.05% taker out, plus real funding

Why it is in the regime section, not the library

It failed the six‑check bar. It is listed under the regime bar, which asks four things of one of the two test windows: a net profit after fees and funding, still a profit without the best trade, a result above 95% of 400 random‑entry runs with the same exit, and at least 20 trades. It clears all four on BNB in both windows: from August 2018 to August 2022 (489 trades) and again from August 2022 to August 2026 (307 trades). On BTC and ETH it clears them from August 2018 to August 2022 only (381 and 483 trades). It is listed as a record, not as a strategy to adopt: the next section says why.

Where it fails the full bar:

• Cross‑asset: BTC and ETH lost from August 2022 to August 2026 (−14.6%, −3.5%).
• Drawdown: 21.5 summed points on BNB from August 2018 to August 2022, against a bar of 20 — its only miss on BNB. BTC's windows drew down 23.1 and 22.8, ETH's 38.6 and 32.4.

It came out of sweep #4, pre‑registered on 2026‑08‑06 to test two corrections to the earlier low‑timeframe sweeps: perpetual fees, and resting limit entries at the maker fee. The sweep killed it, noting that the idea had worked from 2018 to 2022: BTC and ETH lost from 2022 to 2026, and BNB breached the drawdown bar in its first window. It ran before the random‑entry null joined the bar. The re‑audit of sweeps #1–#10, pre‑registered on 2026‑10‑14, ran the null on every window test from those sweeps that made money, made money without its best trade and had at least 20 trades: thirty in all. With a 95th‑percentile bar, about one window test in twenty clears by luck alone. Of the thirty, one or two would by chance; eight did, and four of them are this strategy: BTC, ETH and BNB from 2018 to 2022, and BNB again from 2022 to 2026. A fifth is the same signal bought at market, sweep #1's H3; it is recorded below, not listed.

Why the app can't run it

The app's alerts fire when a bar closes with the entry rules true, and you act at market. This strategy's result comes from how it enters: a limit order resting 0.5% below the signal bar's close for the next two bars, filled only if price trades through it, at the maker fee. About half of the signals never fill. The alerts don't place or describe limit orders yet, and the market version is a different record: bought at market at 0.1% a side, the same signal lost money on BTC and ETH in both windows.

The setup

Market: crypto BNBUSDT perpetual (BTCUSDT and ETHUSDT in 2018–22), 1h trading timeframe. Longs only. Signal (1h, at the close): close > SMA(200) AND RSI(2) < 10. Entry: a limit order 0.5% below the signal bar's close, valid for the next two bars. It fills only if a bar trades below it (a touch is not a fill), at the limit or at a lower open. A new signal replaces an unfilled order. Exit: at the close of the first bar that closes above SMA(5). No stop. Fees: 0.02% maker on the entry, 0.05% taker on the exit (Binance USDT‑M) · funding charged and credited at every settlement held through.

Evidence: two windows on Binance spot candles

BNB, the primary instrument:

Returns are per‑trade percentages summed, like every return in the research log. The price path is Binance spot candles, as in every August sweep; the funding is the real USDT‑M perpetual series. Holding BNB made +1937.8% and +104.3% over the same windows. Where two nulls ran, the table shows the lower percentile.

• Two nulls. The limit fills about half the time, at random bars as at its own signals, so the pre‑registered null — one random signal per real trade — under‑traded: a median of 288 trades against BNB's 489 in 2018–22, and 114 against 307 in 2022–26. An amendment written before any 400‑draw run added a second null that scales the random signals until its median trade count matches the real one: 830 signals for 478 trades, and 827 for 295. A window clears only if it clears both.
• 2018–22: the pre‑registered null had a median of +8.3% and a 95th percentile of +68.4%; the trade‑matched one +19.7% and +81.5%. The real +179.5% sits at the 100th percentile of both.
• 2022–26: pre‑registered median +0.6%, 95th percentile +18.9%, real at the 99.8th; trade‑matched median +0.5%, 95th percentile +31.7%, real at the 95.8th.
• Fees and fills: a fill needs a bar to trade below the limit, not touch it. No fee‑stress run exists for the limit version; the market version is under "What the result depends on".

BTC. Holding it made +203.6% and +169.0% over the same windows:

ETH. Holding it made +288.8% and +10.3%:

• BTC, 2018–22: pre‑registered null median +3.8%, 95th percentile +38.8%, real at the 99.5th (the null took 181 trades to its 381); trade‑matched, from 802 signals, median +8.5%, 95th percentile +57.1%, real at the 95.5th.
• ETH, 2018–22: pre‑registered median +1.6%, 95th percentile +52.7%, real at the 100th (275 trades to its 483); trade‑matched, from 848 signals, median +1.2%, 95th percentile +64.9%, real at the 99.8th.
• 2022–26 on BTC and ETH was not null‑tested: both lost money.

The record by year

On BNB, with BTC's and ETH's returns beneath it:

*2018 from August, 2026 through July. Each trade counts in the year it was entered.

• Most of BNB's money came from 2019 to 2021: +184.9% on 406 trades. Since 2022 it has made +35.1% on 361, and since January 2025 it has lost 2.5 points on 99. No quarter carried it: the best, January to March 2021, made +41.7%.
• Last high: November 2025, at +222.1% summed since August 2018. It ended July 2026 at +211.9%, 10.2 points below. BTC's last high was in March 2022, ETH's in October 2021.
• Last twelve months (August 2025 to July 2026): 60 trades, −2.3%.

What the result depends on

How long the limit rests. The log never recorded it. Its replication section ("Three conventions the log never wrote down") found the August run's numbers only with the limit valid for two bars, the setting listed here; the replication script, fixed before the run, read the log as one bar. The two‑, three‑ and four‑bar runs are disclosed post‑hoc sensitivity runs:

BNB made money in both windows at every setting. BTC lost from 2022 at every setting; ETH made +4.1% at one bar and lost at two to four.

How it enters. Sweep #1's H3 is the same signal bought at the signal bar's close at 0.1% a side, and sweep #4's motivating arithmetic re‑ran H3 with the fees switched off (funding still charged):

At market it traded 718 to 784 times a window, and the 0.2% round trip on each cost more than the signal earned in five windows of six. The limit takes fewer entries, at a better price and the maker fee.

H3's one profitable window, BNB from 2018 to 2022, also clears its null, at the 100th percentile, but against a null that fees distort: its exit, a close above the 5‑bar average, is usually already true at a random bar, so most random entries leave on the next bar and pay the round trip (null median −110.3%, 95th percentile −43.0%). H3's signal is this entry's, and bought at market it lost in five windows of six, so H3 is recorded here, inside this entry, rather than listed on its own.

Before you rely on it

• Most of its money came from 2019 to 2021: +184.9% of +211.9% on BNB. Since 2022 it has made +35.1%, and since January 2025 it has lost 2.5 points.
• The result belongs to the order type. Bought at market at 0.1% a side, the same signal lost money in five windows of six.
• The two‑bar limit is the August run's setting, recovered after the run; at the one bar the replication pre‑registered, BNB made +187.8% and +35.7%.
• On BTC and ETH it stopped working in 2022, and BNB's 2018–22 drawdown, 21.5 summed points, is over the library's 20.

Every number here can be re‑run: npx tsx scripts/research/replicate-sweep4-corrected-costs.ts --phase p2sens --out scripts/research/results/replication-2026-10/sweep4-p2-sensitivity.json for the listed two‑bar runs and every trade (the record by year sums them), npx tsx scripts/research/replicate-sweep4-corrected-costs.ts for the one‑bar runs and the fees‑off autopsy, npx tsx scripts/research/replicate-sweep1-lowtf.ts for H3, and npx tsx scripts/research/null-reaudit-sweeps-1-10.ts --only "#4 P2" for both nulls (--only "#1 H3" for H3's).