jev-nyotti-action (ACTION_V7) — educational experiment

A per-state gradient-boosting policy that imitates the first execution action in the next 15 minutes (hold / open_long / open_short / add / reduce / close) of a user-supplied BitMEX XBTUSD execution export (2018-03 → 2020-06 train). It is paired with a deterministic, direction-neutral decision layer (24h trend filter, 32-bar minimum hold, 8-bar cooldown, no add/reduce). It carries its own paper position and serves jev-nyotti. This is not a Qwen model or an LLM.

Pre-registered gate on Binance 15m BTC/ETH/SOL/XRP, 2022-04-02 → 12-31, a window no model saw before: passed.

4-coin mean, fee-inclusive (7.5 bps/unit)	Net	Fees
V7	+19.0 %	15.0 %
V6 model without rules	−29.2 %	162.7 %
V4	−333.0 %	354.0 %
1-unit buy & hold	−70.2 %	—

Caveats:

• 2022 was a bear year, and the model leans short.
• On 2024, a bull year, the same rules made −1.7 %.
• Leaderboard window 2025-01 → 2026-06: −15.9 %, against buy-and-hold −48.9 %.
• The last 30 days were quiet: 10 coins −0.6 %, and BTC made no entries.

It has not beaten buy-and-hold in bull markets. Treat it as a research artifact, not a strategy.

• Artifact: numeric_policy.json, SHA-256 95531fcda384c095101051bf689947f46177a3a97e47806d3d16bb25f157d284.
• Loaders: inference/jev_inference/numeric_policy.py and decision_rules.py.
• Limits: executed fills only; the source traded as a maker with rebates, while the paper fills are taker; source attribution unverified.
• Not financial advice. No order execution.