truffle-net (v3)

Neural-net microservice in the UVIX mesh (slug: nn). Predicts the probability that BTC's return ~15 minutes ahead is positive. Primary (walk-forward winner): bars. The v3 microstructure model was evaluated on real D1 history and deliberately NOT shipped — see its card.

Model cards (walk-forward, expanding window)

tick model — POST /predict

Task: P(BTC return 180 ticks (~15m at 5s cadence) ahead > 0)
Architecture: MLP 11->16 (tanh) ->8 (tanh) ->1 (sigmoid)
Features: r12, r36, r60, r180, r360, r720, vol180, vol720, vol_ratio, mom_z, range180
Trained through: 2026-10-22T00:10:09+00:00 (n=679 samples, non-overlapping labels)
Walk-forward pooled accuracy: 50.00% vs pooled always-majority base rate 52.70% (pooled n=408). No edge is claimed unless the pooled number beats the base rate — see TRAINING.md.

fold	n_train	n_test	accuracy	majority base	config (validation-selected)
1	271	102	50.00%	50.98%	200ep / L2 0.1
2	373	102	54.90%	52.94%	100ep / L2 0.1
3	475	102	49.02%	54.90%	100ep / L2 0.1
4	577	102	46.08%	51.96%	100ep / L2 1

bars model — POST /predict/bars

Task: P(BTC return 3 bars (15m) ahead > 0) on 5m bars
Architecture: MLP 13->16 (tanh) ->8 (tanh) ->1 (sigmoid)
Features: ret1, ret3, ret6, ret12, ret24, vol12, vol48, vol_ratio, rsi14, range3, vol_z, tod_sin, tod_cos
Trained through: 2026-10-22T04:30:00+00:00 (n=17523 samples, non-overlapping labels)
Walk-forward pooled accuracy: 50.67% vs pooled always-majority base rate 50.43% (pooled n=10514). No edge is claimed unless the pooled number beats the base rate — see TRAINING.md.

fold	n_train	n_test	accuracy	majority base	config (validation-selected)
1	7009	2628	50.53%	50.42%	200ep / L2 0.1
2	9637	2629	49.94%	50.40%	100ep / L2 0.3
3	12266	2628	50.42%	50.38%	200ep / L2 0.1
4	14894	2629	51.77%	50.51%	100ep / L2 0.3

v3 microstructure model — evaluated, NOT shipped (no /predict/v3)

Task: P(BTC Coinbase spot return 15m ahead > 0) from settlement-microstructure features
Architecture: MLP 14->16 (tanh) ->8 (tanh) ->1 (sigmoid)
Features: basis_bps, basis_d15_bps, disp_bps, disp_range_bps, fund_bybit, fund_hl, pred_fund_bin, oi_chg_15m, oi_chg_60m, mark_basis_bps, ls_imb, r15, r60, vol60
Source: wecryp-market-data D1 (spot_bars_1m + deriv_metrics), pulled 2026-10-22 — 19.3 days (2026-10-02 → 2026-10-22), n=1590 samples, 15m stride (non-overlapping labels), up-rate 51.51%
Walk-forward pooled accuracy: 50.73% vs pooled always-majority base rate 51.05% (pooled n=954) — diff -0.32 pts, inside the 95% noise band ±3.17 pts.
Verdict: no edge, and worse — every fold's model predicted "up" on ≥97.5% of its test samples. Out-of-sample the rich-feature MLP collapsed to the majority classifier (the degeneracy rejected at selection time in v1). Shipping it would put a constant predictor behind an API, so the endpoint was not shipped. The pipeline (D1 pull → features → walk-forward) is proven and the rich collector now accumulates this exact feature family at 60s cadence for a retest at greater depth — see TRAINING.md.

fold	n_train	n_test	accuracy	majority base	up-pred frac	config (validation-selected)
1	636	238	51.68%	51.68%	100.0%	100ep / L2 0.3
2	874	239	51.05%	51.88%	97.5%	100ep / L2 1
3	1113	238	49.58%	50.00%	99.6%	100ep / L2 0.1
4	1351	239	50.63%	50.63%	100.0%	100ep / L2 1

Endpoints

GET /health
POST /predict — JSON {"prices":[oldest..newest]}, needs >= 721 prices (~5s cadence ticks)
POST /predict/bars — JSON {"bars":[[ts,low,high,open,close,volume],..]} (Coinbase 5m format), needs >= 49 bars
POST /predict/v3 — not shipped (404): the v3 microstructure model degenerated to the majority classifier in walk-forward; see the v3 card above