← Works

EPHEMERIS

No.
AD-02
Type
Instrument
Status
ACTIVE
Begun
2026

Probability

Five minutes, one question: will Bitcoin close above the line?

Brownian paths against a strike. The counted odds converge on the closed-form ones.

Status

ACTIVE. Research in progress. No claim of edge has been made.

Purpose

Prediction markets run 5-minute BTC binary contracts: will BTC close above a strike price in 300 seconds? The time to resolution is fixed, so pricing is a bounded diffusion problem.

Ephemeris tests one hypothesis: that market participants misprice how remaining time constrains the probability of an outcome. This hypothesis has not been validated.

Architecture

  • Probability model: Brownian bridge formulation solved with the Fokker-Planck equation, compared against market-implied probability.
  • Signals: order flow microstructure (OFI, Kyle's lambda), order book Shannon entropy, Hurst and Lyapunov exponents.
  • Regime classifier: a multi-feature classifier that separates trending, mean-reverting and chaotic conditions.
  • Runtime: a session manager with P&L tracking and loss limits, and a state machine for each market slot.
  • Data flow: the order book feed and a BTC spot feed are recorded separately for replay. Trades and signal snapshots are logged.

Signal thresholds, entry gate logic, classifier inputs and ordering, and runtime design are not disclosed.

Validation

  • Calibration metrics: CRPS and PIT uniformity, with gates specified before testing.
  • Baseline: a coin-flip forecast (p = 0.50). The gate requires CRPS below 0.2125, Sharpe above 2.0 and win rate above 52.5% at p < 0.05.
  • Statistical gate: at least 500 resolved trades before any claim of edge.
  • Simulation: paper trading against live order book data with a small simulated wallet and conservative friction and fill-latency assumptions.
  • Signal distributions: calibrated from 516,000 ticks across 11 slots. These are descriptive statistics, not evidence of predictive power.

On 106 resolved trades, the note reports a 59.4% win rate against the 50% baseline. The PIT histogram shows severe miscalibration. 106 trades is below the 500-trade gate.

Limitations

  • The central hypothesis is untested. No empirical validation has been completed.
  • The same note says both "zero clean trades under correct operating conditions" and 106 resolved trades. That discrepancy is not yet resolved.
  • Infrastructure failures blocked clean execution: a price feed that wasn't recorded separately, a signal gate deployed before it was calibrated, gates disabled during testing, and feed instability.
  • Entry gates do not condition on regime. Regime-stratified calibration is pending.
  • Simulated friction is symmetric. Real friction varies with time of day and liquidity.
  • Thresholds tuned on the same data that generated the signals risk overfitting.
  • The engine this work derives from made "a few hundred dollars" with more capital before its author stopped. That is the known ceiling for this approach.

Current state

Phase 1: paper trading simulator (note dated 18 April 2026). Outcome-conditional signal analysis waits on the 500-trade gate.

Source: Ephemeris: Time-Dependent Probability Estimation in Bounded Binary Markets.