Closes #015 - Deploy Multi-Model Ensemble & Walk-Forward Radar
This commit is contained in:
17
DEV_LOG.md
17
DEV_LOG.md
@@ -140,3 +140,20 @@ This document tracks all modifications, npm packages, active compilation states,
|
||||
### Active Bugs / Compile Status
|
||||
* **Active Bugs**: None.
|
||||
* **Type Checker Status**: Verified clean compilation (`npx tsc --noEmit` returns exit code 0).
|
||||
|
||||
---
|
||||
|
||||
## [2026-06-13] - Multi-Model Ensemble & Walk-Forward Radar (#ISSUE-015)
|
||||
|
||||
### Added
|
||||
* **Stationary Preprocessing & Training Pipeline (Python)**: Created [backend/core/pipeline.py](file:///c:/Users/jannr/.gemini/antigravity/scratch/investment-sandbox/backend/core/pipeline.py) implementing a stationary feature pipeline (Log-Returns, Rolling Volatility, RSI, EMA/SMA distance, spread) on a 365-day rolling window with horizon-cutoff safeguards to prevent look-ahead target leakage. Fits 5 estimators (Random Forest, XGBoost/GB, ElasticNet Logistic Regression, SVM, MLP) for T+1, T+5, and T+10 targets, exporting to `public/data/ensemble_predictions.json`.
|
||||
* **Ensemble API Route**: Deployed [/api/crypto/ensemble/route.ts](file:///c:/Users/jannr/.gemini/antigravity/scratch/investment-sandbox/app/api/crypto/ensemble/route.ts) which parses predictions from the static JSON payload, with a high-fidelity simulated fallback if the file is missing or invalid.
|
||||
* **Walk-Forward Ensemble Radar Grid**: Refactored [CryptoDemo.tsx](file:///c:/Users/jannr/.gemini/antigravity/scratch/investment-sandbox/components/modules/crypto/CryptoDemo.tsx) to mount a glassmorphic Walk-Forward Ensemble Radar dashboard panel displaying predictions, successes/failures, and Expected Value calculations (\(E[\theta] = \alpha / (\alpha + \beta)\)) across 15 independent trackers (5 models \(\times\) 3 horizons) saved in `localStorage`.
|
||||
* **Beta-Posterior Calibration Simulator & Feedback Loop**: Configured the client-side background evaluation loop to resolve forecasts by comparing all 15 estimators against actual returns. Added a calibration simulator to simulate correct/false outcomes across all trackers.
|
||||
|
||||
### Modified
|
||||
* **`QUANT_ROADMAP.md`**: Appended Section IV.V.5 detailing Walk-Forward Validation, Horizon Cutoffs, ML fleet, and multi-tracker conjugate updating.
|
||||
|
||||
### Active Bugs / Compile Status
|
||||
* **Active Bugs**: None.
|
||||
* **Type Checker Status**: Verified clean compilation (`npx tsc --noEmit` returns exit code 0).
|
||||
|
||||
Reference in New Issue
Block a user