Calibration Log: Every Rule Change We've Made, Dated
How our desk grades and delivers signals changes when the numbers make us. Every change since July 2026, what forced it, and what it means for your alerts.
Rules change around here when the numbers make us. This is every change we've made to how our desk grades and delivers signals, when we made it, what forced our hand, and what it means for the alerts you get. When we bench a strategy it keeps running in the dark so it can earn its way back. And we never touch the history.
How this works
Three habits, so that benching a strategy is a decision you can check on, not a loser that quietly went missing.
A benched strategy keeps firing and getting graded where nobody can see it. It comes back on numbers it earns in the dark, not on a hunch.
Before we bench something, we write down what it takes to come back. Usually that's 20 closed signals that add up to a profit. So we can't move the goalposts later.
Whatever we sent you stays in the book exactly as it was graded. The 28 sweeps we delivered are still in there at −8.45R. So is every other loser.
October 2026
Treasury Liquidity launches, as a live trial
Our first strategy that reads how much cash is parked at the Fed instead of a chart. The US Treasury keeps its cash in an account at the Federal Reserve, and money funds park cash at the Fed overnight too. When those two piles shrink fast, that money flows back into the banks, and stocks have tended to rise over the following week. Each day we read both figures from the Treasury's and the New York Fed's public daily reports. When they have fallen unusually fast over the last week of reports, we buy the Dow, the Nasdaq 100 and the S&P 500 at the 4pm New York close, with the stop two average daily ranges away, and we're out at the close five weekdays later (1pm on a half day, and the next session's close if the market's shut). One trade per index at a time.
- On 2016–2026 it made +0.14R a trade over 429 trades, trading costs included, with a profit factor of 1.61 and nine winning years out of eleven. It beat buying on random days at the same point in the month in 50 of 50 tries, so it isn't just the start-of-month effect.
- The catch, and it's why this is a trial and not a full launch: over 2007–2015 it was roughly flat, +0.02R a trade. That missed one of our pre-registered checks. Our best guess is that before 2015 the Treasury kept very little cash at the Fed, so there was little to flow back, but we only worked that out after seeing the numbers, so it doesn't count.
- Before launch we made it stricter: it only reads a Treasury report once that report is actually out, and it never plans an exit for a day the market's shut. Over the same 2016–2026 years that version made the same +0.14R a trade, over 420 trades.
- The live record is the test. If the first 40 trades don't add up to a profit, we retire it and say so here.
Roughly once a month, usually on all three indices the same evening, each trade lasting a week. It's switched on for everyone whose Settings include the 15-minute chart and the Dow, Nasdaq 100 or S&P 500. The trading bots take it at their normal size, in slots of its own, so it doesn't take another strategy's slots. We'll report back here either way.
Rate Decision Drift gets one exception to the dead zone
The Reserve Bank of New Zealand announces at 2pm Auckland time. While New Zealand is on summer time, that puts our Rate Decision Drift entry at 2am UTC (9 or 10pm New York), right inside the dead zone, so those alerts would have been graded but never sent. Rate Decision Drift now skips the dead zone. Every other forex strategy still goes dark from midnight to 3am UTC.
- The dead zone was set from ordinary chart setups. Rate Decision Drift runs on a central bank's clock instead, and its forward test only counts if we actually send the alerts.
- The next four RBNZ decisions land there: October 28, December 9, February 10 and March 17. Winter decisions arrive an hour later, after the dead zone ends.
If you follow NZD/USD on the 15-minute chart and the first hour moves enough, the next RBNZ alert comes at about 10pm New York on October 27.
Rate Decision Drift launches, as a forward test
Our first new forex strategy since the summer, and the first one built around a calendar instead of a chart. Seven central banks each set the rate for one of the majors we cover: the ECB for EUR/USD, the Bank of England for GBP/USD, the Bank of Japan for USD/JPY, the Reserve Bank of Australia for AUD/USD, the Bank of Canada for USD/CAD, the Swiss National Bank for USD/CHF and the Reserve Bank of New Zealand for NZD/USD. On each decision day we wait one hour after the announcement, and if the pair has moved at least 0.15% in that hour we trade in that direction, with the stop one average daily range away. We're out at 4pm New York the next session, whatever has happened. The idea is simple: the first hour's verdict on a rate decision tends to carry into the next day.
- We set the thresholds on 2016–2021 only (183 trades, +0.13R a trade, profit factor 1.50), then ran the locked rule once on 2022–2026, years it had never seen: 162 trades, +0.08R a trade, profit factor 1.27, trading costs included, and it beat random entries on the same pairs in 50 of 50 tries. Over 2005–2015 it made +0.09R a trade on 413 trades.
- The catch, and it's why this is a forward test and not a launch: on 2022–2026 it missed one of our nine pre-registered checks. We ask for a 90% chance that the edge is real, and it scored 89%. A miss is a miss, and we'd rather say so.
- The live record is the test. If the first 40 trades don't add up to a profit, we retire it and say so here.
About three alerts a month, one pair at a time, each lasting until the next afternoon in New York. It's switched on for everyone whose Settings include the 15-minute chart and one of the seven pairs. The next decision days are the Bank of Canada and the RBNZ on October 28, the ECB on October 29 and the Bank of Japan on October 30. The trading bots take it at their normal size. We'll report back here either way.
Pivot Extreme Fade goes on the bench. All of it.
Every Pivot Extreme Fade signal now goes internal, on every chart. Nothing reaches you, and the trading bots stop taking it. It keeps firing and getting graded in the dark, so it can earn its way back.
- On September 25 we benched two slices and kept the rest, and we wrote down that if what was left couldn't add up to a profit over its next 20 signals, the whole strategy would go. We didn't wait for 20. What we kept went 1 for 5, −3.78R, and the fifteen-minute chart, the one slice that had been working, lost all four of its trades this week.
- Counting only the fades we actually sent you, it's 17 wins from 53, −8.11R since it launched in July: +6.84R in July, then a loss every month since.
- A correction while we're here. Our September 25 entry said 53 closed; five of those were graded internal and never sent, so the record you could see stood at 48. The numbers above count only what you received.
30 closed signals in the dark that average at least +0.15R each, judged chart by chart, with the fifteen-minute chart on its own. If it gets there, the comeback gets logged here.
No more Pivot Extreme Fade alerts, whatever your Settings say. On the web it shows as Under review and can't be switched on. Every fade we sent stays in the book, graded as it was, and any fade already open runs to its own stop or target.
Index Snapback reaches seven global indices
Index Snapback now trades the DAX, FTSE 100, CAC 40, Euro Stoxx 50, SMI, Nikkei 225 and ASX 200, with one extra condition. We only buy a foreign index's dip on a day the S&P 500 has a Snapback setup of its own. A foreign index that's oversold by itself, while New York is calm, gets no trade. Everything else is the Snapback you know: the stop three average ranges below the entry, the target back at the index's 5-day average, out within ten trading days, one open trade per index. We don't apply the 200-day filter abroad, because the rule we tested doesn't have one.
- This started as a side note. We tested the plain idea first, buying every one of the seven on a US Snapback day, and it failed. But the breakdown we had written down beforehand showed the whole edge sat in the indices that were oversold too. A foreign index on its own dips, with no US dip, was flat in an earlier test.
- So we wrote that exact rule down and ran it on 2005–2015, a stretch we had never looked at for it: 310 trades at +0.082R each, a profit factor of 1.45, 70% winners, and it beat random days in 50 of 50 tries. It also beat foreign dips that had no US dip behind them, 50 of 50.
- 2016–2026: 313 trades at +0.125R, profit factor 1.84. Six of the seven made money in both decades; the Nikkei lost a little in the first.
- Entering at the foreign market's own open the next morning, instead of right after the US close, still passed. About 28 trades a year across the seven, on the same evenings as the US Snapback entries.
If your Settings include Index Snapback and the daily chart, the seven new markets are in your instrument list as of today. They fire in bunches: a global fear day can send several at once, and a quiet month sends none. The record starts today, wins and losses included.
Fed Minutes launches, as a live trial
Three weeks after every Fed decision, the Fed publishes the minutes of that meeting at 2pm New York. On those mornings we buy the Dow, Nasdaq 100, S&P 500 and Russell 2000 at the 9:30am open and sell at 1:45pm, before the release. The stop sits one average daily range away, and we're never holding overnight.
- Over 2016–2026, trading costs included: +0.106R a trade over 320 trades, a profit factor of 2.22, and it beat random mornings in 50 of 50 tries.
- We then ran the identical rule, untouched, on eight foreign stock indices it had never seen. All eight made money over 2016–2026 too.
- The catch: over 2005–2015 it lost a little in the US (−0.03R a trade) and broke even abroad, and 2022–2024 were small losing years too. So the edge looks real since 2016 and weak or absent before it, and we can't tell you why. A sibling test on the Fed's other 2pm report, the Beige Book, found nothing. The live record is the test. If the first 24 trades don't add up to a profit, we retire it and say so here.
About eight mornings a year, up to four alerts at once. The first one is due Wednesday, November 18. It's switched on for everyone whose Settings include the 15-minute chart and a US index. We'll report back here either way.
Holiday Eve launches, as a live trial
Another strategy built around the calendar. Markets tend to drift up into a holiday, so two sessions before each US market holiday we buy the Dow, Nasdaq 100, S&P 500 and Russell 2000 at the 4pm New York close, as long as the index is above its 200-day average. We sell at the close of the last session before the holiday (1pm when the market closes early). The stop sits one average daily range away, and we never hold over the holiday itself. If News Night is already trading that evening, Holiday Eve sits it out.
- Full disclosure first: this idea already failed once. On October 1 we tested it without the trend filter, and it flunked its final check on 2025–2026, making almost nothing (+0.01R a trade). We retired it that day.
- We came back with one change, picked from six we wrote down before looking: only buy when the index is in an uptrend, the same filter Index Snapback and Overnight Reversal use. With it, the rule made money in all three stretches we have: +0.118R a trade on 2005–2015, +0.064R on 2016–2024 and +0.048R on 2025–2026.
- But we chose that filter after seeing all of those years, so none of it counts as a clean test. The live record is the test. If the first 30 trades don't add up to a profit, we retire it and say so here.
About six or seven evenings a year, up to four alerts at once, and fewer when markets are falling. The first possible one is Tuesday, November 24, two days before Thanksgiving. It's switched on for everyone whose Settings include the 15-minute chart and a US index. We'll report back here either way.
News Night adds the Russell 2000
News Night now trades the Russell 2000 (US2000) too, with exactly the same rules: bought at the 4pm New York close the evening before a Fed decision, jobs report or CPI release, out at the 9:30am open. US2000 still runs nothing else but Index Snapback and News Night.
- We ran the unchanged rules on Russell history they had never seen, trading costs included. 2016–2026: 326 trades at +0.042R each, a profit factor of 1.29, and it beat random nights in 50 of 50 tries. 2005–2015 held up too: 340 trades at +0.034R.
- These are the same evenings as the other three indices, so it adds about 30 alerts a year, not new nights.
On a news night you can now get up to four News Night alerts at once. If your Settings include US2000 and the 15-minute chart, you'll get the Russell ones too.
News Night launches, as a live trial
A new strategy built around the calendar. The evening before a Fed decision, the monthly jobs report or the CPI inflation report, we buy the Dow, Nasdaq 100 and S&P 500 at the 4pm New York close and sell at the 9:30am open. The jobs and CPI reports come out at 8:30am, so they land inside the trade. On Fed nights we're out hours before the 2pm decision. The stop sits one average daily range away. There's no chart pattern here: the date is the signal. Monday to Thursday only, so never over a weekend.
- Over 2016–2026, trading costs included: Fed nights made +0.096R a trade over 255 trades, and jobs and CPI nights +0.038R over 744. Both beat random nights in 50 of 50 tries.
- Small per trade, but steady: 9 of the 11 years made money (2016 and 2022 lost), across about 30 nights a year.
- Our first test failed: Fed nights narrowly, jobs and CPI nights clearly. It charged trading costs at 2005–2015 index prices, when the same spread cost several times more per trade than it does today. At today's costs both kinds of night make money in both decades and pass our full test on 2016–2026. We picked that costing, and that window, after seeing the first result, so the clean test is the one that starts now.
About 30 evenings a year, usually three alerts at once, one per index. The first one is due Tuesday, October 13, the evening before CPI. It's switched on for everyone whose Settings include the 15-minute chart and US30, US100 or US500. We'll report back here either way.
Index Snapback adds the Russell 2000
Index Snapback buys a hard dip in a US index that's still in an uptrend, for a bounce back to its 5-day average. Until today it watched the Dow, Nasdaq 100 and S&P 500. Now it watches the Russell 2000 (US2000) too, with exactly the same rules. Nothing else runs on US2000: we tested two more of our index strategies there, and both failed.
- Our first test, on 2016–2026, wasn't convincing: it made money, but it didn't beat random entries often enough. So we ran the unchanged rules on 2005–2015 Russell history it had never seen, trading costs included: 90 trades at +0.105R each, a profit factor of 1.77, and it beat random entries in all 50 of 50 tries.
- 2016–2026 still made money: 77 trades at +0.065R each.
- That's 7 or 8 more Snapback alerts a year, roughly 30% more than before.
Most US2000 alerts land on the same sell-off days as the other three indices, so a rough day can now bring four Snapback alerts at once. If you already get Snapback alerts, we've added US2000 to your markets; you can remove it in Settings.
VIX Inversion launches, as a live trial
A new strategy for fear spikes. The VIX is Wall Street's fear gauge. When it closes at 25 or higher and above its own 3-month version, short-term fear is running hot, and we buy the Dow, Nasdaq 100 or S&P 500, priced at the daily close, for a bounce back to the index's 5-day average. The stop sits three average daily ranges away, and the trade is out within 10 trading days. When the VIX is under 25 we still track the signal, but we don't send it.
- Over 16 years of history, trading costs included: about +0.08R a trade, about 11 trades a year on average but in bursts (17 a year over 2025–26), and a profit in 11 of the 13 years it traded.
- Only 2022–24 would have cleared our full test, at +0.26R a trade. It lost in 2010, and in 2020, the COVID crash and the biggest panic in our data: −4.5R over 39 trades.
- We drew the VIX-25 line after seeing every year, so the only clean test is the one that starts now.
About 11 alerts a year, in bursts and often on all three indices at once, when fear spikes. It's switched on for everyone: you'll get it if your Settings include the daily chart and US30, US100 or US500. Treat it as a live trial: we'll report back here either way.
September 2026
Four-hour alerts were arriving three hours late. Fixed.
We were checking for new four-hour candles on the wrong clock. The market draws those candles from 5pm New York time, and we looked an hour before each one closed, so we didn't see it until the next look, three hours later (two in winter). Every four-hour alert reached you that late. From today we check the moment each candle closes, so the alert arrives within seconds.
- Over the last two weeks, four-hour alerts went out a median 3 hours after their candle closed. Fifteen-minute and one-hour alerts go out in about 10 to 25 seconds.
- One of the six four-hour candles each day, the one that closes at 5pm New York, was never checked at all. It is now, so expect a few more four-hour alerts. Same strategies, same rules: just a candle we'd been skipping. It arrives about 11 minutes after it closes, because 5pm is when spreads blow out and we don't send anything until they settle.
- Daily alerts had the same problem: they now arrive about 12 minutes after the daily close instead of about 3 hours.
- The book prices every four-hour alert at its candle's close. Until today you got those alerts hours after that price, so the four-hour results in the book assume an entry you couldn't actually have taken. We're not re-grading them. We're telling you so you can weigh them.
Four-hour alerts now arrive when the candle closes, and there will be a few more of them. Nothing already in the book changes.
Two slices of Pivot Extreme Fade go on the bench. The rest stays.
Pivot Extreme Fade on the one-hour chart now goes internal, all of it. So does the four-hour chart whenever the touch isn't a forex pair inside the 1–1.5×ATR pocket, or it's USD/CAD. The fifteen-minute chart and the in-pocket four-hour fades keep coming to you exactly as before.
- Two months of live fades, 53 closed. The one-hour chart went 4 wins from 17, −5.7R, and it lost inside the pocket and outside it. The four-hour chart outside the pocket went 3 from 11, −3.1R.
- The two slices we're benching add up to 28 signals, −8.8R, one win in four. What's left adds up to 25 signals, +5.4R, winning 44% of the time.
- The surprise is which chart works. The backtest said four-hour was the one; live, the fifteen-minute chart is the only timeframe in profit (7 of 16, +5.0R). We're not promoting it on 16 trades. We're just not benching it.
- The test we set on August 21 — twenty premium signals, then judge — can't run: the premium rule has fired once in five weeks. So the test moves to what's still delivered.
The benched slices come back the usual way: 20 closed signals in the dark that add up to a profit. And we've written down the other direction too. If what's still delivered can't add up to a profit over its next 20 closed signals, the whole strategy goes on the bench.
Fewer Pivot Extreme Fade alerts — none from the one-hour chart, and four-hour ones only when the setup is the kind that has actually paid. Everything you've already been sent stays in the book, graded as it was. And from today a signal keeps the grade it got the moment it fired; we don't re-grade it later, even when the rules change.
Liquidity Sweep goes on the bench. All of it.
Every sweep signal now goes internal, and no score, reward or session rule can rescue one. Nothing reaches you until we've rebuilt it.
- In August we let forex sweeps back in on a short leash and wrote down the test they'd have to pass. They failed it: 0 for 3, −3.00R, every one a full stop. The last one (GBP/JPY, M15) opened and got stopped the morning we pulled the plug.
- All-time, the sweeps we sent you: 4 wins out of 28, −8.45R. Last 90 days: 2 of 18, −6.38R.
- We were tempted by a fix: all six winners ever came between 11:15 and 15:30 UTC. But fitting a three-hour window to six trades is exactly the kind of storytelling this page exists to stop.
20 closed sweeps in the dark that add up to a profit, or a change to the entry itself.
No sweep alerts for now. The 28 you got stay in the book, graded as they were.
If the stop sits inside the spread, it isn't a trade
Some signals had a stop closer than twice the instrument's usual spread. You can't actually execute that, and the R on it is make-believe. Those now go internal. We bumped the engine version to v3 so anyone can see where the line is.
- 60 days back, that class was 6 signals, 4 of them sent to you. All 6 lost, full stop. So this rule takes away nothing you'd have wanted.
- We also tried a stricter version at five times the spread and threw it out: it would have cut about 29% of your alerts, including 15 winners.
Nothing good disappears. A kind of alert that could only lose is gone.
August 2026
EMA 200 Reclaim on the new markets earns its way out of the dark
When we added metals, oil and indices in July, we ran their M15 reclaim signals in the dark to see if they actually worked. They did. They now go out like everything else.
- Aug 14 to 26, in the dark: 17 closed, 47% winners, +21.95R. About +1.29R a signal.
- We replayed our normal filters over those 17. The over-extension rule alone would have caught the six losers (−6R), and the reward floor two more. Hiding all of them was overkill.
- What gets through: 6 premium (+12.68R) and 3 watchlist (+13.52R). US500 on H4 stays out. It went 0 for 7.
More reclaim alerts on non-forex M15, same rules as everything else.
Weekend Gap Fill couldn't fire. Fixed.
We launched Gap Fill on Aug 17 and it was silent every Sunday after. Not a quiet market and not a rule blocking it. A bug: the detector threw away the Sunday reopen bar as "not finished yet" before it could see the gap. Fixed on the 25th. The first live signal fired Sunday Aug 30 at 21:15 UTC, NZD/USD short on M15.
Sunday-evening gap alerts actually arrive now. Judge the strategy on a small live count until it grows.
M15 breakouts: the New York rule goes, the reward floor drops to 2R
Three things. We dropped July's rule that only let M15 breakouts fire during New York hours. We lowered the reward floor for M15 breakouts from 3R to 2R, while everything else keeps 3R. And M15 breakouts that open outside 9 to 18 ET get labelled watchlist, not premium, until live results earn the upgrade. Went live Aug 26.
- We replayed two windows, Jun 8 to Aug 21 and March through May, charging the spread and assuming the worst on exits. The New York rule's logic didn't hold in either: signals outside those hours made money both times (+16.1R over 12 recently, +8.5R over 23 in spring), and the 9 to 12 ET overlap was actually the worst slot.
- The 3R floor was blocking M15 longs that planned under 3R (+11.0R over 19, profit factor 2.14) while waving through the 3 to 4R ones, which went 0 for 7. A 2R floor beat both 3R and no floor at all.
when 20 live off-hours signals add up to a profit.
More M15 breakout alerts. Some say watchlist instead of premium for now.
Pivot Extreme Fade: premium has to be earned
When we launched pivot fades we labelled every one premium. That stopped. Premium now takes all five: forex, the H4 chart, the first touch of that level this period, a stretch of 1 to 1.5× ATR, and not USD/CAD. Everything else still goes out, labelled watchlist.
- The week of Aug 17 went 1 for 14, −11R, all of them labelled premium. Lifetime, the live fades win 30.4% of the time. A fixed-2R strategy needs 33.3% just to break even.
- The 20-month backtest (187 trades, profit factor 1.33) makes about 93% of its money inside that 1 to 1.5× ATR pocket: +33.0R over 66, 52% winners, positive in 6 of 9 pairs and in both years.
- USD/CAD is the chronic one: −14.1R over 22, 18% winners. And across 299 trades on every market we track, the edge only exists in forex. It flips negative on metals (−9.5R over 40) and energy (−3.2R over 23).
Fewer premium labels, same alerts. Expect two or three premium fades a month.
Mean Reversion gets graded on the reward it plans, not its score
The score gate is gone. It was backwards: signals scoring 70 to 77 made +0.91R on average, 78 to 84 lost −0.66R, and 85 and up went 0 for 3. Now the label follows the planned reward. Premium at 2R or better, watchlist at 1.4R or better, internal below that. This replaces the flat 3R floor, which a pivot-targeted reversion trade almost never reaches by design.
- 90-day replay: the premium group made +2.9R over 17 (35% winners). The 1.4 to 2R group we now demote lost −2.9R over 11.
- I picked a 2R premium bar over 3R (+6.9R over 14) to get the flow back after a week with zero premium alerts.
Mean-reversion alerts flow again, graded on what they actually plan to make.
M15 breakout shorts go on the bench
M15 breakout shorts now go internal. Shorts on H1 and H4 still go out.
- All-time they're −90.6R over 226 trades, winning 18% of the time, and they lost in every month we checked: April −23.8R over 57 (during our best breakout month ever), May −43.8R over 116, June −12.3R over 31, July −8.8R over 20.
- H1 shorts made +4.5R and H4 shorts +8.7R at 54% winners, positive nearly every month.
20 closed M15 shorts in the dark that add up to a profit.
No M15 breakout short alerts until they earn it back.
Asian Range Breakout goes on probation
Signals that would have been premium are capped at watchlist. Monday signals go internal. The score floor went up to 80, longs need six more points than shorts, and we halved how far past the range a signal is allowed to chase, from 0.30 to 0.15 of the range.
- June +44.0R over 64. July −13.0R over 46. August −20.4R over 24, with one winner. Every pair lost both months, and the tuned version went 0 for 9 in August. That's the market changing, not a knob we can turn.
- Mondays: −18.4R over 25 all-time. Scores 75 to 79 lost −9.9R over 47 while 80 to 84 made +28.0R over 49. Longs lost −15.0R over 90; shorts made +25.6R over 44. Chasing 15 to 25% past the range cost −22.2R over 31.
a rolling 45-day check on the tuned version, written down in advance.
Fewer ARB alerts, and none labelled premium until it earns the label back.
Weekend Gap Fill launches
A new strategy. At the first M15 close after the Sunday reopen, when price has gapped 10 pips or more from Friday's close, we fade it back toward Friday. Stop one gap beyond the open, 48 hours to work. Premium from day one.
- A year across 9 pairs, charging three times the normal spread because Sunday nights are ugly: profit factor 1.50, +0.171R a trade, all 9 pairs positive. Per weekend, profit factor 1.84. Every stop-size and cost combination we tried came out positive.
One or two Sunday-evening alerts a week, when a gap qualifies.
The dead zone: forex signals opening between midnight and 3am UTC go dark
Any forex signal whose bar opens between 00:00 and 02:59 UTC, which is 8 to 11pm New York, after the close and before Tokyo wakes up, now goes internal, whatever the strategy.
- A 120-day study by hour of day: that's the only three-hour band that lost in both halves of the window. Hours 00, 01 and 02 together came to −50.1R over 103 forex signals (leaving ARB out), and 0 for 16 in the last 90 days.
- This is not an "Asia is bad" rule. The same study showed that blocking all of Asia would have cost +27.1R from April to June.
if those hours start paying again. They keep getting graded either way. We owed ourselves a re-check on Sep 4 and haven't done it yet.
No alerts from those three hours.
Internal means internal, in every market
We found a leak. The email digest had a forex-only rule that let internal signals on other markets slip through, about 29 a month. Closed. Now an internal signal reaches nobody, anywhere. The ones that leaked before the fix still count in the record, because you got them.
Same day, two more things. We started the four-weekly rule review this page grew out of. And we let forex sweeps back in on a short leash, watchlist only, while metals and oil sweeps stayed out: 0 for 6, with fantasy reward targets of 5.6 to 14.7R, all stopped. We wrote down the test the forex ones had to pass by Sep 1. They didn't pass it. See Sep 4.
The book only has what we actually sent.
Minimum planned reward goes from 2R to 3R
A signal now has to plan at least 3R of reward for every 1R it risks, or it goes internal, on M15, H1 and H4. This roughly halved premium alerts. Two strategies have since earned their own lower floors: Mean Reversion on Aug 18 and M15 breakouts on Aug 22.
- Over 150 days, the 2 to 3R band we cut lost −13.6R over 127 signals. That supports the rule.
- But month by month it went June +6.0, July −20.4, August +4.9. July did all the work. And the win rate falls as planned reward rises (35% down to 21%), so the idea behind the rule isn't as clean as it sounded. We're saying so.
- Under 2R stays out regardless: −8.8R over 88.
by Sep 15, or once 40 signals have been cut by this rule, whichever comes first. If the 2 to 3R group would have made money, the floor drops to 2.5R.
Fewer, bigger-reward alerts. This one's on watch, not settled.
July 2026
EMA 200 Reclaim: the over-stretched ones stop going out
If the 50 and 200 EMAs are more than 2 ATR apart when the signal fires, it goes internal.
- Jul 8 to 31: signals above that line went 1 for 18, −15.45R. Below it: 30.9% winners, +11.24R. The odds of that split being luck are about 1 in 80, and it holds however you slice it: by pair, by date, by market, by direction, by timeframe.
- The score couldn't see it. It actually hands out maximum points for wide separation, and over the same window score and results had basically no relationship (r = +0.004 across 99 signals).
- Checked again at the Aug 7 review: it blocked 4 signals, all losers. Four for four. Best rule we have.
Fewer reclaim alerts, and the missing ones were the losers.
First bench: Liquidity Sweep. Plus, no M15 breakouts on metals and oil.
Liquidity Sweep (2 wins in 21, −0.59R a trade) went internal instead of getting deleted, which is the habit this whole page runs on. M15 breakouts on gold, silver and oil were blocked outright. And we limited M15 breakouts to New York and overlap hours, a rule that didn't survive contact with the data and came off on Aug 22.
- Metals and oil M15 breakouts: 0 for 8 in the July review, 0 for 9 in June. Our pip settings were built for forex and they don't translate to gold's volatility.
Three sources of losing alerts closed. One rule later reopened on evidence. Two stand.
The scores that predicted nothing
Mean Reversion's score floor came off after we found the score was backwards: higher score, worse results (correlation −0.30). We also rebuilt how it scores "stretch", so an extreme stretch, which usually means a trend day, scores zero instead of top marks. Asian Range Breakout's score floor went up from 62 and 58 to 75.
- ARB under 75: −10.9R over 26. At 75 and up: +37.9R over 55.
- Mean Reversion with the score gate gone: +6.81R in the study window.
Score never gated Mean Reversion again. ARB's floor later went to 80, on Aug 17.
Breakout score floor set at 75
Breakout signals scoring under 75 stopped going out. We set the floor at 70 in the morning and raised it to 75 the same day.
- 45 days of data on the current engine: the 70 to 74 band was the worst breakout group per trade, and 75 and up the only one that made money.
- It cost about 55% of breakout volume. We noted at the time: if alerts get too thin, revisit with fresh numbers before lowering it.
About half as many breakout alerts, from the half that paid.
What we're watching
Open questions with dates on them. When one gets answered, it becomes an entry above.
If the 2 to 3R signals we've been cutting since Aug 3 would have made money by the time 40 of them have piled up, the floor drops to 2.5R.
On 50 clean days, watchlist averaged +0.74R and premium −0.03R. That's upside down. Premium has to beat watchlist or the labels get fixed.
−30.6R over 73 trades at 12% winners, with no rule on it yet. We're looking before we write one.
Lift the midnight-to-3am bench if those hours are paying again. They keep getting graded in the dark either way.
Sweeps and M15 breakout shorts return at 20 closed signals in the dark that add up to a profit. Pivot Extreme Fade at 30, averaging at least +0.15R. Asian Range Breakout on its 45-day rolling check.
Words we use above
- R
One unit of the risk you took at entry. +2R means the trade made twice what it risked. −1R is a full stop.
- Premium · watchlist · internal
Premium is our best-evidenced setup. Watchlist still goes out, at lower conviction. Internal gets graded but never sent to anyone.
- The book
Every signal we actually sent to subscribers, graded as it closed. Internal signals are left out of every public number. It's all here.
- In the dark
A benched strategy keeps firing and getting graded where nobody can see it, against a bar we wrote down on the day we benched it. The opposite of fitting a rule to the past and calling it proof.
Every number on this page comes from the signals we actually sent, graded at close. Internal signals are left out of every published figure. The full statistics update continuously; this log updates the day a rule changes.
Want the changes as they happen, plus every signal graded, every week? The Dossier is free.
We got receipts.
Every signal our desk fires, graded in R — winners and losers, no cherry-picking. One email, every Friday.
Share this post
Know a trader who could use this one? Send it their way.