The Lab

How we tested it

• The same rebuild and the same test as the 4-hour page. Each signal was judged by the price 1, 3, 6, 12 and 24 bars later, in the direction it names, after paying a market order in and out (0.11%). It was compared with 400 draws of as many random bars.
  • "Beats chance" means its average is above 95% of the random draws.
  • "Worse than chance" means it is below 5% of them.
• Under 10 signals is counted as too few to judge. Those tests are listed but left out of the counts.
• Three groups of charts:
  • Bitcoin over two years (11 October 2024 to 11 October 2026), on 11 charts from 5 minutes to a week.
  • Bitcoin from 2020, on the day, 3-day and week charts, run two ways:
    • with open interest, from 1 September 2020;
    • without open interest, from 1 April 2020.
  • Unseen data: the 8- and 12-hour sell signals (S?, D, flush) on data they had not seen: Bitcoin from 2020 to October 2024, and ETH and SOL over the same two years. The rule was fixed before the run: a signal is worth watching only if it beats chance at most horizons there.
• Why luck gives about 5% each side: a signal with no skill is just one more random pick, and a random pick lands above the 95% line 1 time in 20. So of 1,035 judged tests, about 52 would beat chance by luck, and about 52 would do worse.
  • Example, one test on unseen data: S? on ETH's 8-hour chart, real flow, one bar (8 hours) on, fired 20 times and averaged +0.74% after fees.
  • Chance's 400 averages ran from −0.74% at the 5% line to +0.65% at the 95% line.
  • +0.74% is above the 95% line, so it beats chance. It is one of the 3 of ETH's 40 tests that beat chance, where luck gives 2.
  • The 4-hour page works a test through step by step.
• Why a count is a lean, not proof: neighbouring horizons reuse the same signals, so one lucky set of signals can pass at several horizons at once.
• Why under 10 is too few: with a handful of signals, one big move decides the average. Three signals that each caught a sell-off look like a strong signal, but they are three events.
• Every setting was fixed and written down before the first run.

What we found

1. The 8- and 12-hour signs did not hold on unseen data. Out of 125 tests, 9 beat chance and 2 did worse; luck alone gives about 6.
   • On Bitcoin before October 2024, 1 beat chance and 2 did worse, of 50.
   • On ETH, 3 of 40 beat chance, all at the very next bar only.
   • On SOL, 5 of 35 beat chance, scattered across horizons.
   • No signal beat chance at most horizons, so none is named worth watching.
2. Fast charts (5 to 30 minutes) sit far outside its design. There, the fee is the whole story. On 5 minutes, every signal averaged about −0.11% (the fee) at every horizon up to 2 hours: the moves were smaller than the cost of a trade. On 5 and 15 minutes, 29 tests did worse than chance and 1 beat it, of 180.
3. The 1- and 2-hour charts showed nothing beyond luck (7 beat chance and 5 did worse, of 180). On the 4-hour chart, its design, this two-year run gave 3 and 6 of 90: the same as the 4-hour page.
4. Slow charts give too few signals.
   • The 3-day and week charts gave under 10 signals each over two years.
   • From 2020, the week chart still gave too few.
   • The 3-day chart from 2020 had 3 and 4 beats of 30.
5. The day chart from 2020 leaned slightly to the sell side. It had 7 beats of 75 with open interest and 9 of 60 without, none worse, where luck gives 3 or 4. But the lean belongs to the early years:
   • S? a day later made +4.0%, +1.6% and +1.5% in the years to October 2023.
   • Then it made −1.1%, +0.2% and −0.2%.

It was also found after looking, so it would need unseen data before it meant anything.

1. Overall, 67 of 1,035 judged tests beat chance and 49 did worse. Luck alone gives about 52 on each side.

The 8- and 12-hour row on Bitcoin over two years is the same data where the signs were first seen, so it is not new evidence. The unseen-data row is the test.

Limits

• The tests overlap: neighbouring horizons share the same signals, so the counts against luck are a lean, not proof.
• One rebuild of the indicator, from its Pine source. Its author's own reading of a signal may differ.
  • Pine (Pine Script) is TradingView's own language for chart indicators. An indicator's Pine source is its code, which anyone can read on TradingView. For example, ta.sma(close, 20) is Pine for the average of the last 20 closes.
• Prices are Binance spot closes. There is no slippage and no funding cost on the trades, and wicks inside a 5-minute candle aren't seen.
  • Slippage is the gap between the price you expect and the price a market order fills at: the order takes whatever the order book offers. For example, with half a coin offered at 100 and the rest at 100.20, a market buy of one coin averages 100.10, 0.1% of slippage.
  • A wick is the thin line above or below a candle's body: how far the price went during the candle before coming back. For example, a candle that opens at 100, dips to 95 and closes at 99 has a wick down to 95.
• The coins move together, so ETH and SOL are not fully independent of Bitcoin.
• Leaving out the 2 June 2026 sell-off lowered the 8-hour flush results (+2.7% to +2.1% eight days on) but did not flip them.

Update, 27 October 2026: where the idea comes from

Added after this page was published. The text above is left as it was.

• The indicator's own ideas (spot pressure, funding, the random-bar test) and where each comes from are on the 4-hour page, in its update of the same date.
• Testing on data a rule was never fitted to is the out-of-sample test. Mervyn Stone set out the method of holding data back to judge a prediction in 1974. Here the unseen data was Bitcoin before October 2024, and ETH and SOL.
• Many tests, some pass by luck. The classic guard is Carlo Bonferroni's (1936): divide the 5% bar by the number of tests.
  • For example: 1,035 tests at 5% each would need each test to clear 5% ÷ 1,035, about 0.005%.
  • 400 random draws can only measure down to 1 in 400, which is 0.25%. So this page counts passes against the number luck gives (about 52) instead.
  • Why every test needs a random twin is on our theory page "Hard-to-beat markets, and why every test needs a random twin".
• Theirs and ours. The indicator and its signals are its author's. Stone's held-back data, Bonferroni's guard and the Monte Carlo method are credited below. The choice of charts and coins, the rule fixed before the unseen run and the "too few" cut at 10 signals are ours.

Credits

• SvL v4.5 (BTC Spot vs Leverage), credited at the top of this page, published on TradingView and shown here with permission. The indicator is theirs; we tested it on other charts and coins.
• Nicholas Metropolis and Stanisław Ulam, "The Monte Carlo Method", Journal of the American Statistical Association, 1949: drawing random comparisons many times.
• M. Stone, "Cross-Validatory Choice and Assessment of Statistical Predictions", Journal of the Royal Statistical Society, Series B, 1974: judging a rule on data held back.
• Carlo E. Bonferroni, "Teoria statistica delle classi e calcolo delle probabilità", Pubblicazioni del R. Istituto Superiore di Scienze Economiche e Commerciali di Firenze, 1936: the guard for many tests.
• The test design follows our page "A TradingView indicator, rebuilt and replayed on two years of Bitcoin".
• Data: the exchanges' public candles, funding, open interest and taker records named on this page.
• PerpGremlin and Claude, The SvL on other charts and coins, TerminalGremlin, 2026: the tests, the code and the words.