Provyan Research · run on 16 Nov 2026
Two things we tested and did not add: funding rates and 4-hour candles
Both ideas were tested the way every rule here is: fixed on past data, then measured on data that came after. Neither has earned a place in the live rule. The figures are below so the decision can be checked.
1. Funding rate as a filter
87 of the 94 coins have a USDT perpetual on Binance. For each, the funding rate known at a daily close was taken as the average 8-hour rate over the last 7 days, read from the exchange's public archive. The usual walk-forward was then run: studies recomputed up to a cutoff, calls of the following year collected, three years in all. Each call was labelled by its funding, and four labels fixed in advance were compared with the unlabelled calls on the first two years (16 Nov 2023 to 15 Nov 2025) and read off the third (16 Nov 2025 to 8 Nov 2026).
| Down calls | Calls | Right | Average after costs |
|---|---|---|---|
| All calls with a funding reading | 247 / 211 | 62% / 67% | −0.2% / −0.5% |
| Funding negative | 56 / 107 | 75% / 65% | +7.0% / −3.2% |
| Funding at 0.03% per 8 hours or more | 18 / 0 | 50% / – | −25.5% / – |
| Funding in the coin's own lowest tenth | 48 / 51 | 83% / 71% | +5.1% / −2.2% |
| Funding in the coin's own highest tenth | 47 / 14 | 45% / 71% | −10.8% / +6.3% |
| Up calls | Calls | Right | Average after costs |
|---|---|---|---|
| All calls with a funding reading | 328 / 303 | 52% / 45% | +4.7% / −0.3% |
| Funding negative | 46 / 139 | 52% / 46% | +2.0% / +2.0% |
| Funding at 0.03% per 8 hours or more | 45 / 2 | 67% / 100% | +7.9% / +53.8% |
| Funding in the coin's own lowest tenth | 46 / 63 | 57% / 56% | +2.9% / +1.5% |
| Funding in the coin's own highest tenth | 101 / 39 | 61% / 46% | +11.5% / +4.7% |
Each cell reads “first two years / third year”. Calls with a proof score of 55 or more; “right” means the price moved the called way over 7 days; averages are per call after 0.3% costs. Only calls on coins with a funding reading that day are counted, which is why the totals are smaller than in the ledger's replay.
Decision: not added. The bar, set beforehand, was at least 60 calls in each period, 5 points more often right than the same calls unlabelled, and a better average, in both periods. No label passed. The nearest was down calls made while funding sat in the coin's own lowest tenth: right 83% of the time on 48 calls, then 71% on 51, with an average of +5.1% and then −2.2%. Too few calls, and the second average is the wrong side of zero.
2. The daily method on 4-hour candles
Every detector was run on 4-hour candles exactly as it runs on daily ones, so a “day” in each rule became one candle: a 14-candle RSI, a 200-candle average, a 10% drop within one candle. 94 coins, four years of candles. Studies were scored on candles up to a cutoff with the usual luck test (10,000 random draws, corrected across all studies), and the calls of the following year were checked 4, 12 and 24 hours later. Two cutoffs a year apart give two test years: 15 Nov 2024 to 15 Nov 2025, and 15 Nov 2025 to 15 Nov 2026.
Checked 4 hours later
| Calls | Number | Right | Average after costs | Separate days |
|---|---|---|---|---|
| All calls | 15,374 / 12,117 | 51% / 54% | −0.09% / −0.06% | 366 / 365 |
| Up calls | 8,096 / 6,285 | 51% / 52% | +0.19% / −0.20% | 364 / 364 |
| Down calls | 7,278 / 5,832 | 51% / 56% | −0.41% / +0.08% | 366 / 365 |
| Proof score 70 or more | 12,190 / 8,976 | 52% / 54% | −0.04% / −0.03% | 366 / 365 |
| Up, score 70 or more | 6,668 / 4,765 | 52% / 52% | +0.25% / −0.18% | 361 / 360 |
| Down, score 70 or more | 5,522 / 4,211 | 51% / 57% | −0.40% / +0.13% | 366 / 365 |
Each cell reads “first test year / second test year”. An ordinary candle was followed by a higher price 4 hours later 49% and 46% of the time.
Checked 12 hours later
| Calls | Number | Right | Average after costs | Separate days |
|---|---|---|---|---|
| All calls | 9,436 / 9,964 | 55% / 53% | +0.24% / −0.06% | 366 / 365 |
| Up calls | 5,538 / 5,923 | 55% / 52% | +0.72% / +0.01% | 355 / 364 |
| Down calls | 3,898 / 4,041 | 53% / 56% | −0.44% / −0.16% | 365 / 365 |
| Proof score 70 or more | 6,975 / 7,561 | 55% / 54% | +0.33% / −0.02% | 365 / 365 |
| Up, score 70 or more | 4,252 / 4,727 | 56% / 52% | +0.82% / −0.02% | 337 / 358 |
| Down, score 70 or more | 2,723 / 2,834 | 54% / 57% | −0.45% / −0.02% | 354 / 363 |
Each cell reads “first test year / second test year”. An ordinary candle was followed by a higher price 12 hours later 49% and 46% of the time.
Checked 24 hours later
| Calls | Number | Right | Average after costs | Separate days |
|---|---|---|---|---|
| All calls | 9,230 / 9,908 | 56% / 57% | +0.49% / +0.20% | 366 / 365 |
| Up calls | 5,040 / 5,140 | 56% / 54% | +0.93% / +0.47% | 348 / 363 |
| Down calls | 4,190 / 4,768 | 56% / 60% | −0.04% / −0.09% | 365 / 364 |
| Proof score 70 or more | 6,831 / 7,666 | 56% / 58% | +0.61% / +0.39% | 366 / 365 |
| Up, score 70 or more | 3,671 / 3,964 | 56% / 56% | +1.07% / +0.67% | 336 / 355 |
| Down, score 70 or more | 3,160 / 3,702 | 57% / 61% | +0.07% / +0.09% | 361 / 363 |
Each cell reads “first test year / second test year”. An ordinary candle was followed by a higher price 24 hours later 49% and 45% of the time.
Decision: not built. Across the three checks, calls were right between 51% and 57% of the time, and with 0.3% costs on moves this small the average per call sat close to zero: +0.49% and +0.20% at 24 hours. Up calls checked 24 hours later were the steadier side, at +0.93% and +0.47% a call, right 56% and 54% of the time: too thin a margin, on calls that bunch into the same days, to build a product on. No 4-hour signals are sealed today.
Why publish tests that changed nothing
Because a site that reports only what worked cannot be checked. The daily rule, its replay and its live record are on the ledger page; the full count of what holds up is in the findings report; the method is on the methodology page.
Statistics about the past on Binance data, not advice.