PhaseBreak 2026

Cross-Domain Phase Transition Detection with Log-Periodic Power Law Analysis

Framework for detecting phase transitions across 5 domains using LPPLS (Sornette 2003), HMM regime detection, and Bayesian Survival analysis. Validated on real yfinance, Sentinel-2 satellite, FHFA housing, and commodity data.

Scientific Hypothesis

LPPLS parameters (m, ω) from financial bubbles and geological spectral anomalies are drawn from the same distribution — preliminary evidence of universal phase transition signatures.

Core Math: LPPLS (Sornette 2003)

ln(E[p(t)]) = A + B(tc - t)^m + C(tc - t)^m * cos(ω * ln(tc - t) + φ)

7 parameters: tc (critical time), m (exponent), ω (frequency), A, B, C1, C2

Domains (5) — 58 benchmark episodes

Domain	Data Source	Episodes	v2 Results
Finance	Yahoo Finance	11 bubbles + 9 controls	Precision=78%, Recall=64%
Commodities	Yahoo Finance (futures)	6 bubbles + 4 controls	Precision=67%, Recall=50%
Housing	FHFA + Zillow	10+6 episodes	Precision=67%, Recall=33-50%
Geology	Sentinel-2 temporal series	Seismic precursor patterns	KS p>0.05
Fraud	Transaction timelines	Doomsday Bayesian, C-index +27%	Survival model

Domain Maturity Tiers

Tier	Domains	Meaning
Supported	Finance, Commodities	Validated on real data, forward-tested, production-usable
Experimental	Housing, Epidemics	Works but limited recall, needs more data/tuning
Science-only	Geology, AI Compute, Landslides	Research validation, not for operational decisions
Separate track	Fraud	Different model (survival), not LPPLS bubble detection

Pipeline Architecture (v2)

Layer A: Screening → data quality + HMM regime detection
Layer B: Structural Fit → LPPLS + soft scoring + tc uncertainty + adaptive windows + HMM prior
Layer C: Scientific Inference → cross-domain KS, universality (offline only)

Entry points:

• run_full_pipeline() — v2 recommended path
• run_legacy_pipeline() — backward-compatible path
• HMMLPPLSEnsemble.analyze() — legacy ensemble

Web API & Dashboard (NEW)

FastAPI Server — REST API for LPPLS detection:

pip install fastapi uvicorn[standard]
python -m service.server.main
# API: http://localhost:8000
# Docs: http://localhost:8000/docs

React Dashboard — Visual interface with real-time scanning:

cd service/frontend
npm install && npm start
# Dashboard: http://localhost:3000

API Endpoints:

• POST /api/v1/scan — Scan multiple tickers
• GET /api/v1/scan/{ticker} — Scan single ticker
• GET /api/v1/scorecard — Get current predictions
• GET /api/v1/domains — List available domains
• GET /api/v1/benchmark — Get benchmark summary

See service/README.md for full documentation.

One-Command Replication

# Linux / macOS / Git Bash
bash reproduce.sh

# Windows PowerShell
powershell -ExecutionPolicy Bypass -File reproduce.ps1

Creates a virtual environment, installs all dependencies, runs 268 tests, then reproduces the full v2 benchmark (58 episodes), ablation study, and baseline comparisons. Results saved to data/v2_results/.

Quick Start

pip install -e ".[dev]"
pytest tests/ -v                           # run all tests
python -m src.benchmark.v2_benchmark       # run official v2 benchmark
python -m src.benchmark.v2_ablation        # run ablation study

Key Contributions

1. Multi-window DS LPPLS Confidence Indicator — fit on multiple overlapping windows, consensus = confident signal (Sornette 2015). Predicts tc with 1-19 day accuracy on known bubbles.
2. Tightened Sornette filters — the critical engineering contribution: raises precision from 55% to 100%, eliminates 5/6 false positives.
3. HMM-gated LPPLS ensemble — Hidden Markov Model pre-screens regime → LPPLS fits only in bubble state (novel combination, precision 80%).
4. Doomsday Bayesian survival — Gott (1993) random observer assumption as Cox PH feature for fraud scheme lifetime prediction. +27% C-index on synthetic data (upper bound).
5. Cross-domain parameter comparison — KS tests on (m, ω) between finance and geology: p > 0.05 (cannot reject H0, small sample).

v2 Enhancements (confirmed)

Component	Status	Impact
Soft scoring	Confirmed	Quality score 0-1 instead of binary pass/fail
tc uncertainty (bootstrap)	Confirmed	[p10, p90] intervals, mean width ~4 days
Adaptive windows	Confirmed	Frequency-aware: daily/quarterly/highvol presets
HMM prior weighting	Confirmed (mixed impact)	Reduces compute on normal periods; +1 FP on finance ablation
Pipeline separation	Confirmed	Detector ≠ science inference
Triple split	Confirmed	Train/val/test per domain
Adversarial controls	Confirmed	6/6 correct (TP=1, TN=5)
EW metrics	Confirmed	Lead time, coverage, interval width

Exploratory modules (not in default pipeline):

• Conformal prediction — secondary uncertainty method (needs calibration set)
• Changepoint (CUSUM) — diagnostic only, not integrated into verdict
• Wavelet LPPLS — CWT spectral diagnostics, ω domains incompatible
• EWS (critical slowing down) — weak on real data
• AI council — requires Ollama, heuristic fallback

v2 Ablation (Finance, 12 episodes)

Layer	TP	FP	Precision	Recall
Raw LPPLS	4	0	100%	67%
+ Hard filters	4	0	100%	67%
+ Soft scoring	4	0	100%	67%
+ Adaptive windows	4	0	100%	67%
Full v2 (+ HMM prior)	4	1	80%	67%

Project Status

• LPPLS model + tightened Sornette filters (Stage 1)
• Multi-window confidence indicator (Stage 1)
• HMM-gated ensemble — precision 80%, recall 67% (Stage 1.5)
• Geology — LPPLS on real Sentinel-2 data, KS p>0.05 (Stage 2)
• Fraud survival — Doomsday + Cox on synthetic data (Stage 3)
• Critical Slowing Down — EWS layer (Stage 3.5)
• Adversarial AI council — framework, heuristic mode (Stage 4)
• Cross-domain correlation — KS + Mann-Whitney + bootstrap (Stage 5)
• Paper draft — 7 pages, 3 figures (Stage 6)
• v2 integration — soft scoring, uncertainty, adaptive windows, HMM prior
• v2 benchmark — 58 episodes across 6 categories (finance, commodities, housing, housing_monthly, adversarial, forward)
• v2 ablation — layer-by-layer evaluation
• 5 domains validated (finance, commodities, housing, geology, fraud)
• Forward validation 2024-2025: Nvidia/Nikkei/BTC detected, 0 FP
• 268+ tests passing

License

MIT