V7: rejected before PAPER trading
V7 was designed to test whether a more autonomous learning system could discover a durable trading edge. It never reached PAPER deployment: the offline research gate rejected it first.
Initial hypothesis: reinforcement learning
The first architecture used SAC reinforcement learning across BTC, ETH and SOL. The agent controlled directional exposure while reward was based on net PnL. Multiple iterations added deadbands, turnover penalties and explicit HOLD behavior. The recurring failure was overtrading: weak predictive information was converted into frequent position changes whose costs overwhelmed the signal.
Strategic reboot: edge first
Rather than tuning RL around its losses, we removed RL from the decision loop. V7 was rebuilt as a supervised research pipeline: market data → expected-return forecast → economic edge test → position/risk only if the forecast survived costs. FLAT was the default. No PAPER executor was created.
What we tested
The research used BTC, ETH and SOL with 5-minute market data and prediction horizons from 15 to 120 minutes. Models included fixed Ridge and histogram gradient boosting configurations. Feature experiments covered price/volume structure, trend, volatility, cross-asset relationships and derivatives information such as funding and futures basis.
Validation was chronological rather than random. Expanding walk-forward OOS windows were used, future-return labels were kept outside training boundaries, trading costs were included, and confidence buckets were examined only as diagnostics. We explicitly stopped selecting isolated winners once that risked turning research into strategy fitting.
Final frozen viability test
The final decision test evaluated the derivatives concept as a family rather than choosing the best-looking asset or horizon. At 0.07% assumed trading cost, only 1 of 12 asset/horizon combinations had a positive average net result, and that result was close to zero. None were profitable in all four OOS windows, and none were profitable in at least three of four windows.
Across 48 asset/horizon × OOS-window observations, 6 were positive and 42 were negative. Mean directional hit rate was about 53.5%, showing that some predictive information existed, but mean net edge after costs was approximately −0.0275% per evaluated trade. Raising the confidence threshold did not make the family-level mean net edge positive.
Decision
V7 was rejected before PAPER trading. The key distinction was predictive accuracy versus economic usefulness: the models could detect some directional structure, but not strongly or consistently enough to pay realistic execution costs. Further selection of individual assets, horizons or thresholds would have increased selection bias rather than established robustness.
Why publish a system that never launched?
Pre-deployment rejection is a successful research outcome when the evidence does not justify deployment. Publishing the failed path preserves the methodology, prevents the same idea from being quietly recycled, and makes clear that only strategies that pass the research gate are allowed into forward PAPER observation.
Experiment record
V7 RL / edge-first ML
Version: V7 RL / edge-first ML. Status: REJECTED. Mode: OFFLINE — never deployed to PAPER or LIVE.
Hypothesis: An autonomous learning or supervised forecasting system may discover a durable net trading edge.
Methodology: Chronological walk-forward out-of-sample validation on BTC, ETH and SOL, frozen model configurations and explicit execution costs.
Dates: Exact experiment start and retirement/rejection dates are not published in this record.
Measured results: The preserved frozen test reported 6 positive and 42 negative OOS observations; no PAPER or LIVE result exists.
Conclusion: Rejected before PAPER: weak and unstable economic edge after costs did not justify deployment.
PAPER is simulated execution, not LIVE capital performance. Existing results and historical records are preserved.