BTC Market Research Benchmark

Generated 2026-04-26T08:01:47.793337+00:00 · symbol BTC · 8486 one-minute rows · 168h source · 2705 features · runtime 2103.542s.

Abstract

We evaluate whether saved orderbook sentiment, CVD/footprint flow and per-exchange depth slices contain a measurable predictive edge for BTC direction across horizons from 1 minute to 72 hours. We benchmark majority and momentum baselines against linear (Ridge return regression, balanced logistic), tabular gradient boosting, and a class-balanced 2D market-state CNN. We then conduct repeated-fold blocked Monte Carlo ablations on (a) individual depth ranges, (b) cumulative depth stacks (0.5%; 0.5+1%; 0.5+1+2%; ...), (c) random Monte Carlo combinations of depth ranges, and (d) per-exchange feature subsets, both alone and combined with the global aggregate. The best configurations and laws are reported below; the live system is recommended to deploy the highlighted candidate.

Recommended deployment configuration: lookback_24h_correlation_rule at horizon 8h using feature set lookback_24h. Balanced accuracy 0.9650, raw accuracy 0.9650, edge over majority 0.4055.

Methodology

Each minute is one row. Features are global aggregate sentiment by depth (0.5%-100%), CVD/footprint stats (per market type), price returns and rolling stats, and (when enabled) per-exchange sentiment xsent_<exchange>_<market>_<depth>_<metric>. Targets are 1m-4320m forward log returns and their direction.

1. Classical pass: chronological 70/15/15 split, train on the first 70%, evaluate on the last 15%. Reports majority baseline, 3m momentum rule, single-feature correlation rule, balanced logistic regression, Ridge return regression, and HistGradientBoosting.
2. Repeated blocked Monte Carlo ablations: 8-12 random training-window starts (each window block keeps order intact), per-fold metrics averaged. Avoids one-lucky-split bias.
3. Depth ablations:
   • Individual depth: orderbook_<d>%_only, without_orderbook_<d>% (leave-one-out).
   • Cumulative depth: 0.5; 0.5+1; 0.5+1+2; 0.5+1+2+5; ... add one range at a time.
   • Monte Carlo depth combos: random subsets of size 2-5 such as 0.5+2+5%.
4. Per-exchange ablations: each exchange alone, each + global, each + global + CVD, all-exchanges union, all + global. We use minute-aggregated orderbook_sentiment rows filtered by asset='BTC' for exchange IDs 9, 15, 16, 17, 18, 19, 21, 22, 23.
5. Image CNN: 32-row windows pivoted to a 2D feature image, class-balanced cross-entropy, AdamW, validation early stopping. Treated as a smoke test until repeated.
6. Scoring: balanced accuracy (primary; immune to class imbalance), raw accuracy, edge over majority baseline, predicted vs actual up rate. A result is only treated as a candidate law when balanced accuracy ≥ 0.60 and edge over majority > 0.

Dataset

Leaderboard (Classical Pass)

Ranked by balanced accuracy with edge over majority and raw accuracy as tiebreakers.

Single-Depth Ablation

8-12 blocked Monte Carlo folds. Each row tests whether one specific depth slice (or its absence) carries the signal.

Heatmap

ablation_ridge

ablation_logit

Cumulative Depth Stacks

Adds one depth at a time: 0.5% alone, then 0.5+1%, then 0.5+1+2%, ... This isolates whether deeper book layers strictly add information.

Heatmap

cum_depth_ridge

cum_depth_logit

Monte Carlo Depth Combinations

Random subsets of depth ranges, e.g. 0.5+2+5%, to discover combos no human-prior would test.

Heatmap

mc_depth_ridge

mc_depth_logit

Per-Exchange Edge

Each exchange tested in isolation (only that venue's sentiment) and combined with the global aggregate. Ranking averages balanced accuracy across all ablation horizons.

Full per-exchange ablation table

Heatmap

exchange_ridge

exchange_logit

Per-Individual-Market Edge

Each individual venue+symbol+market_type triple tested in isolation, then combined with the global aggregate, and with global + CVD. Surfaces venues like BTCFDUSD on Binance where market-makers and arbitrage bots leave a strong signal.

Full per-market ablation table