memescalp — memecoin prediction experiment

Prediction skill only — nothing is bought or sold, real or simulated. Claude calls UP/DOWN vs a coin-flip control; accuracy is scored against actual prices.
experiment running for

Pass / Fail — prediction skill, nothing else

Scoreboard

Claude accuracy
Classical math
ML (learning)
Hedge (RCH)
Random accuracy
Coin-flip baseline
50%
Pending horizons

Running accuracy (%) — resolved predictions

■ Claude  ■ Classical math (OU 1930 / variance-ratio 1988 / EMA)  ■ Random control  · dashed line = 50% coin flip

Cumulative hypothetical return (%, frictionless)

Sum of signed moves if each call were followed with a fixed stake — an upper bound; real trading costs would subtract ~2-3% per round trip.

Capital growth — ₹2,000 ($22.99) compounded through every resolved call

■ Claude  ■ Classical math  ■ Random control  · dashed line = starting capital · frictionless (no fees/slippage) — an upper bound on what following each arm would have grown to.

Kelly-sized capital — betting only when a trailing edge exists

Half-Kelly fraction from each arm's trailing 30-call win rate (Kelly 1956), capped at 25% of capital, zero until 10 outcomes exist. A flat line means Kelly's verdict is "no edge — don't bet". This is the fastest honest compounding possible for whatever edge is real.

Calibration — does "60% confident" mean 60%? (Brier 1950)

Claude Brier score
Math Brier score
Uninformative baseline
0.250
Lower Brier = better. Rows compare stated confidence with actual hit rate — calibrated confidence is what makes Kelly sizing valid.

Measured earn-rate — what each arm would generate in a real up/down market (1% fee, half-Kelly)

Computed from each arm's last ≤200 scored calls at its live call rate. "Bankroll for $1/min" inverts your target: the capital needed at CURRENT measured skill to earn $60/hour. Below-breakeven skill = no bankroll size works; the fix is skill, not size.

Live hot board — best coin right now

Re-ranked every 5 seconds from raw feed ticks (the finest resolution public data allows): live score = recent movement × tick persistence − round-trip cost − shallow-pool penalty. This is the analyzer's recommendation of what's most active/tradable this moment; the forecaster sees the same numbers.

Prediction ledger

Token catalog — trending candidates ranked by metrics

EvoML lab notebook — self-experiments, one generation every 30 min

Each row is one tournament: the champion genome (model kind, sensory chromosome size, invented genes, call threshold, self-mutation rate) and its purged-holdout score. Invented genes are expressions the model built from its own inputs and kept because the lineage carrying them survived.

Claude's evolving playbook — self-written lessons, updated every 30 min

no playbook yet — written after 20 resolved calls

Latest Claude prediction (verbatim)