or4k2l/THUNBIT
Demand-regime instability detection for daily SKU demand series with STABLE / DRIFT / SHIFT state outputs.
What's novel
THUNBIT is an experimental detector for demand-regime instability in daily SKU-level demand series. It is not a forecasting model, not production-ready, and has not been formally validated on real-world data. In Cooperation whith AI.
Code Analysis
11 files read · 5 roundsDetects regime changes (STABLE/DRIFT/SHIFT) in daily SKU-level demand series by combining KS-test, variance, and autocorrelation evidence into a confidence score, then applying a calibrated state machine with baseline-normalized scoring to reduce false alerts.
Strengths
Exceptionally honest documentation that explicitly states limitations (unsolved false-alert problem, synthetic-only validation, manual parameter calibration), combined with a well-structured progressive refinement chain (V4→V4.5) that clearly documents trade-offs at each iteration. Input validation is thorough and the state machine logic is clean and well-documented.
Weaknesses
The core statistical methods (KS test, variance ratio, ACF change) are standard and the combination is straightforward; the real innovation is in the calibration/normalization layer which is still acknowledged as unsolved. All validation is on synthetic data with manually tuned parameters, and the six detector variants share significant code duplication in their rolling detection loops rather than composing shared logic.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
31
Contributors
2
Files
20
Active weeks
1
Repository
Language
Python
Stars
0
Forks
0
License
Apache-2.0