Bitcoin Astro Predictor | Version 2

Statistical Evidence Brief for TP-Segment Regime Alignment

This brief is a compact scientific-style summary derived from RESEARCH/grid_search_massive_label_weighted_statistical_significance_tp_segment_weighted.ipynb and top-100 grid-search significance results. The objective is to test whether regime-segment alignment between astro-only model outputs and Bitcoin market structure is better than random.

Core data source: data/market/reports/turning_massive_label_grid_top100_tp_segment_weighted_g1p5_mind5_tail1_pred-hard_label.csv
Each candidate was evaluated against a permutation null (median permutations per candidate: 451).

Executive Summary

Analyzed Candidates

Candidates With p < 0.05

Positive Effect Size (Δ > 0)

Combined Fisher p-value

Latest Market Date

Active Model eval_id

Interpretation: aggregate evidence is inconsistent with a purely random alignment model. The result supports the V2 statement: "weighted TP-segment alignment is non-random at aggregate level."

Methods (Condensed)

For each top candidate from the massive turning-point grid search, we compute a weighted segment-hit statistic and compare it to a permutation null. The null preserves structure but randomizes alignment, producing an empirical p-value: p_shift_weighted_hit.

We then aggregate evidence across candidates using: (1) sign test on effect direction (delta_weighted_hit_vs_null > 0), (2) Fisher combined p-value, (3) distributional plots and model-quality trade-off review.

Key Visual Evidence

Histogram of weighted-hit p-values — **Figure 1.** Distribution of permutation p-values for top-100 candidates. The left tail contains more low-p candidates than expected for a purely random ranking.

Observed vs null weighted-hit rate scatter — **Figure 2.** Observed weighted-hit rate vs null mean. Most points are above diagonal; best candidate (eval_id=7279) is a strong positive outlier.

**Figure 3.** Effect size distribution for `delta_weighted_hit_vs_null`. Mean effect is positive (0.078), bootstrap 95% CI [0.039, 0.118].

Top 10 models effect sizes — **Figure 4.** Top-10 candidates after significance-first sorting. Green bars denote p < 0.05.

Top 15 tradeoff chart — **Figure 5.** Recall-min / MCC / objective-profit trade-off among top-15. Marker size encodes `test_profit_y_obj`.

Primary Metrics Table

Metric	Value	Interpretation
p_shift_weighted_hit < 0.05	--	Concentration of low p-values is above random baseline expectation.
p_shift_weighted_hit < 0.01	--	Several candidates stay significant under stricter threshold.
Δ weighted-hit > 0	--	Directional tilt to positive effect size, not 50/50.
Sign test p-value	--	Probability of 65+ positives out of 100 under random direction is very low.
Fisher combined p-value	--	Aggregate evidence across candidate set is strongly non-random.
Stouffer combined z / p	--	Independent aggregate test confirms positive global shift.
Mean Δ (bootstrap CI95)	--	Effect-size center remains above zero under resampling.

Best V2 Candidate Snapshot (eval_id=7279)

Metric	Value
p_shift_weighted_hit	--
delta_weighted_hit_vs_null	--
weighted_hit_rate_obs / null_mean	--
test_profit_y_obj	--
test_recall_min	--
test_recall_gap	--
test_mcc	--

Conclusion

At aggregate level, the V2 TP-segment metric shows statistically significant non-random alignment.
The strongest candidates present substantial positive effect size relative to permutation null.
Quality metrics (recall balance and MCC) still require continuous monitoring alongside objective profit.

Verdict (V2): aggregate evidence supports non-random TP-segment alignment between astro-only model signals and Bitcoin regime structure.