First read of our own btc15/eth15 observe data (332 windows each, 5 days): the fair-value model has strong signal on 15m, but the market ask ≈ the actual win rate across the entire τ curve — so there is no tradeable regime. Early NO-GO; keep the observe container banking data. This is the population read the 15m tape verdict predicted (signal is real, mechanically dead on the population).

For Agents — binding rules

  1. 15m signal ≠ 15m edge. The market is efficient on 15m — the ask tracks the actual win rate across the whole τ curve. Do NOT build a 15m money lane.
  2. Keep the btc15/eth15 observe container running (banking data). The only “maybe” is the early-τ pocket, and it is currently execution-impractical (below).

The signal is real (model works on 15m)

From 332 btc15 + 332 eth15 observe windows (5 days):

  • p_fair is monotonic in outcome (calibration shape is sane).
  • Conviction windows (p_fair ≥ 0.80) resolve up 98% actual — high-conviction is well-founded.
  • Under-confident tails, just like 5m.

So the fair-value model transfers to the 15m interval. The signal is not the problem.

The economics kill it (market is efficient)

The market ask ≈ actual win rate across the ENTIRE τ curve:

  • ask 0.95–0.98 vs actual 0.98–1.0;
  • real_edge ≈ ±1% everywhere — no bin where the ask meaningfully lags the outcome.

There is no τ pocket where the ask lags reality — unlike 5m, where the [60,150)s window is exactly where the oracle-lag ask lags. On 15m the market has already priced the lag out across the board.

The two bigger apparent edges are unreliable

Apparent edgeWhy it’s not usable
<120sexecution-impractical (too little time to fire/fill through pmv2’s path)
600–720sn=290, thin, and contradicts prior tape (the competitor 15m tape)

Verdict

15m signal ≠ edge; the market is efficient. Don’t build a 15m money lane. Keep the observe container banking data — the early-τ pocket is the only “maybe,” and it is currently execution-impractical. This closes the 15m question from our own data, consistent with the competitor-tape verdict that any 15m capture needs a fresh calibration model AND a real τ pocket the population does not offer.