🌱 Survival score track record

Which signals actually mattered

Each step counts only coins still alive at the previous check (1h → 24h → 7d → 30d); coins that haven’t reached a check yet are left out, not counted as alive or dead. “Survived” is the chance of making it to the next check when the signal was on, shrunk toward the average as if we had 10 extra average coins, so small samples look neutral. Older outcomes count less (half weight after 45 days). Lift above 1 = survives more often than average.

Exploration slice

About 20% of coins (picked by a fixed hash of the address, so a coin never switches) also get a hidden score from the original untuned weights, plus their raw numbers are logged, so learning can’t lock in on whatever worked lately. Split so far: pre_explore 2 · main 41.

• Main (shown score): 0 coins with a 7-day outcome · no data yet · not enough data (need 200)
• Explore (untuned score): 0 coins with a 7-day outcome · no data yet · not enough data (need 200)

Brier score: lower is better (0 = perfect, 0.25 = coin flip).

Models tested side by side

Same coins, same signals, same re-checks. Only Model A is shown on coin pages; the others are scored silently until one clearly wins with at least 200 coins per check. Outcome = still alive at that check (coins that haven’t reached it are left out). Brier and log loss: lower is better. AUC: higher is better (0.5 = guessing).

• A: Fixed-step creep from hand-set weights (public score)
• B: Bayesian nudge (shrinkage, ridge centred on current weights, hazard, decay)
• C: Plain logistic hazard regression (one coefficient per signal)
• D: Gradient-boosted trees (trained on the box)
• E: Survival / time-to-peak hazard model (trained on the box)
• F: Price time-series forecaster (Chronos, box-run; 1h/24h/7d only, needs price history)
• explore: Untuned v1 weights on the 20% exploration slice

Exploration slice vs main, 7-day Brier per model: A main – (0) / explore – (0) · B main – (0) / explore – (0) · C main – (0) / explore – (0) · D main – (0) / explore – (0) · E main – (0) / explore – (0)

Model E time-to-death concordance: no data yet · not enough data

• B · 2026-10-28 06:09 UTC · Need 60 coins (intake) with a 7-day outcome, have 0.
• A · 2026-10-28 06:09 UTC · Need 60 coins with a 7-day outcome, have 0.

How the weights learn

• Model A (public): once 60 coins have a 7-day outcome, signals seen in at least 30 move ±1 point per update when their survival lift is above 1.15 or below 0.85 (max 6 signals, ±10 from the original, 1-point floor).
• Model B starts once 60 coins from our automatic intake of new launches (not visitor lookups) have a 7-day outcome; at most once every 20 hours.
• Fits all signals together (ridge logistic regression on the step-by-step survival data), pulled toward the current weights rather than zero, so each version is only a nudge.
• A signal moves only if it fired on at least 30 resolved coins with both survivors and deaths, and 90% of 40 resampled fits agree on the direction.
• At most ±2 points per signal and 12 points total per version. Weak signals shrink to a 1-point floor but are never removed or flipped.
• Tested on the newest 30% of coins after training on older ones; used only if Brier score and log loss are no worse. Every version is kept and can be rolled back.

Current weights: v1 (hand-set, not tuned yet)

The public model (A) changes by small fixed steps once 60 coins have a 7-day outcome (see learning rules above). A full logistic refit is only tried manually and only after 200 outcomes; it switches only if it beats the current weights on a held-out 30%.