The Lab

• PROVEN its chances beat the simple up-rate on calls it never saw, by more than luck, corrected for the fact that we test 15 patterns at 3 horizons, so at most about 1 in 10 proven patterns is a fluke
• ON TRIAL no verdict yet: too few calls, or an edge still inside luck
• FAILED SO FAR its chances scored worse than simply using the all-coin up-rate
• RETIRED switched off; kept on the board for the record
• Hit rate counts calls that leaned one way; the lazy call always bets the side most coins ended on. Skill compares the chances themselves with the all-coin up-rate. Only proven patterns vote on a coin.

Setups all setups →

Event patterns all patterns →

150 checked102 held48 flipped

share of cases higher the next day, older cases → newer cases · last column = cases

1. Oct 1, 2026patternsCHANGEDsynthetic test with planted effects

   Trust bar on a market with nothing planted

   tried

   Score all fifteen patterns at 1, 7 and 30 days (45 records) on a synthetic 600-coin market of random walks with a shared Bitcoin factor and market-wide drift spells, where no pattern effect was planted.

   result

   With one 2-standard-error test per record, 1 of the 45 passed by luck: Volatility Squeeze at 30 days (1,227 calls, 2.02 standard errors), and it then cast the only vote on 25 coins.

   now

   Proven now also has to survive a fluke filter across every check in the build (Benjamini-Hochberg at a 10% false discovery rate over all 45 records, with 2 standard errors kept as a floor): at most about 1 in 10 proven patterns is a fluke. On the 600-coin market nothing passes: a pattern with no edge posts a record as strong as Volatility Squeeze's 2.2% of the time, and the best of 45 checks has to be under 0.22%; all 600 coins read 'no proven pattern'. On the demo market 5 of 45 records stay proven (Lookalike, the event pattern and Rubber Band at 30 days, Rubber Band at 7 days, and Breakout at 30 days at 1.6% against a 1.8% cutoff); Market Breadth (2.0 and 2.05 standard errors) falls back to on trial.

2. Oct 1, 2026patternsKEPTdemo data

   Cases counted once per episode

   tried

   Count a pattern's past cases once per episode, on the first day its condition turns on, instead of once for every day it stays on.

   result

   On the demo market the ten market patterns were on for 32,236 coin-days in 4,221 separate episodes. A 50/200-day cross stays on for 20 days (1,462 coin-days, 77 crosses) and the 90-day trend state for weeks (14,750 coin-days, 865 episodes), so counting days would have counted the same move many times.

   now

   The cases behind each line and in the playground are episodes; a condition that switches off for less than 5 days and back on is the same episode. Each pattern keeps its 5,000 most recent cases.

3. Oct 1, 2026patternsCHANGEDdemo data

   One vote per horizon

   tried

   Let each pattern vote at 1, 7 and 30 days only where its own record at that horizon is proven, instead of one 30-day vote.

   result

   On the demo market 16 of 21 coins lean higher at 30 days and 5 come out even. At 7 days only Rubber Band has earned a vote: 1 coin leans higher, 20 come out even. At 1 day no pattern has, so all 21 read 'no proven pattern'.

   now

   The vote is stored per horizon. A pattern proven at 30 days has no say at 1 day until its 1-day record earns it.

4. Oct 1, 2026patternsFAILEDsynthetic test with planted effects

   Planted next-day breakout pop

   tried

   To give the new 1-day odds something real to find, add a push the day after a breakout in the demo generator: half a typical daily move up after a close above the 30-day high, down after a close below the 30-day low.

   result

   Across the demo's 2.5 years of prices the push showed: 65% of breakouts closed higher the next day and 36% of breakdowns, against 52% for all coin-days. In the scored year it did not: its 150 breakouts closed higher 52% of the time, the same as all coins, and the 1-day record sat 0.04 standard errors from zero. A smaller push (0.3 of a daily move) also knocked the 30-day records of Breakout, Market Breadth and Rubber Band below proven.

   now

   Removed. The demo keeps its two planted effects (unlock drift and rubber band) with unchanged prices.

5. Oct 1, 2026patternsKEPTdemo data

   Eight trader patterns added

   tried

   Add eight patterns traders already know: golden / death cross, breakout of the 30-day high or low, capitulation (down 25% in a week), green / red streaks, fresh all-time highs, deep drawdowns, volatility squeezes and market breadth. Nothing was planted for any of them in the demo.

   result

   Breakout earned proven at 30 days on the demo (168 calls, 70% right, 2.2 standard errors) and still clears the stricter fluke filter adopted the same day. It rides the demo's planted rubber band: breakouts above the 30-day high ended higher 30 days later only 34% of the time and breakdowns 75%, against 57% for all coins, and most breakouts are stretched prices. Market Breadth beat the up-rate too (248 calls, 2.0 standard errors; 2.05 at 7 days) but not by enough to survive 45 checks, so it is on trial. The other six stay on trial with 10 to 86 scored 30-day calls each, under the 100 needed.

   now

   Fifteen patterns are scored the same way at 1, 7 and 30 days. A market pattern draws a line only on the days its condition is on.

6. Oct 1, 2026patternsKEPTdemo data

   Next-day odds for every pattern

   tried

   Give every pattern a 1-day chance next to its 7- and 30-day ones, scored on its own 1-day calls and held to the same bar: beat the all-coin up-rate on calls it never saw, by more than luck.

   result

   No pattern cleared the bar at 1 day on the demo market. The closest stayed inside luck: Breakout (116 calls, 1.3 standard errors above zero), the event pattern (256 calls, 1.2) and Rubber Band (180 calls, 1.2); Lookalike made only 13 one-day calls. A single day's move is mostly the whole market's move.

   now

   Every line shows its 1-day chance and its own 1-day record. At 1 day every pattern is on trial, and the 1-day vote reads 'no proven pattern' on all 21 demo coins.

7. Oct 1, 2026scoringRETIREDproduct decision

   One score per coin (the Slate)

   tried

   Rank every coin by one composite score built from setup, risk, record, dilution and pulse.

   result

   Retired as a separate board when the site went prediction-first.

   now

   The old board redirects to Predictions, where coins are ranked by their chance of ending higher.

8. Oct 1, 2026calculatorREPLACEDdemo data

   Trade replay on the closest matches only

   tried

   Replay a leveraged trade only on the few past cases closest to the coin's situation.

   result

   Its chance of profit and the prediction above it disagreed by 10 points on JUP (59% vs 49%).

   now

   The calculator replays the same weighted pool of cases the prediction uses, so the two agree (59.2% vs 59.1%).

9. Oct 1, 2026patternsCHANGEDsynthetic test with planted effects

   Proven = beats the up-rate

   tried

   Mark a pattern proven once its calibrated chances beat the simple all-coin up-rate over 100 or more calls.

   result

   Two patterns with no real effect in the test market (Last Year This Time, Bitcoin Shadow) passed by luck: 20 coins that move together over one year give only about 12 independent months at 30 days.

   now

   Proven also needs the weekly gains to sit at least 2 standard errors above zero; both noise patterns now stay on trial.

10. Oct 1, 2026patternsREPLACEDsynthetic test with planted effects

    One shrink factor per pattern

    tried

    Pull each pattern's chance toward the all-coin up-rate by a single factor measured on its own past calls.

    result

    In a test market with a planted unlock effect (41% chance higher over 30 days) it flattened the unlock pattern to 51%: each event line rests on only about 13 cases, so the noise dragged the factor down.

    now

    Event patterns are calibrated in groups (event kind and size); the same test now reads 43%.

11. Oct 1, 2026patternsREPLACEDdemo data

    Raw pattern chances

    tried

    Show each pattern's raw chance of ending higher next to its line.

    result

    The 90-day trend pattern said 78% higher while its 30-day calls had been right 49.5% of the time.

    now

    Every chance is calibrated against the pattern's own record, and a pattern without a proven record shows no number at all.