Jann/investment-sandbox
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Closes #MASTER-RESET-024 - Implement layout, performance stats, and quant ML integration

Browse Source
This commit is contained in:
Antigravity Agent
2026-06-17 18:38:54 +02:00
parent 776a41f159
commit 71b7683370
11 changed files with 5469 additions and 458 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
QUANT_ROADMAP.md
View File

@@ -41,6 +41,9 @@ This document serves as the permanent, centralized system architecture design an
* **Phase 8.0: Isolated PEAD Screener & Fallback Warning Graphic**
* *Features*: Integrated zero-coupling Post-Earnings Announcement Drift (PEAD) Screener parsing reported vs. consensus EPS surprise vectors. Implemented an `isLiveApi` flag to detect and handle offline API or rate-limit simulated fallbacks. Mounted glassmorphic warning status badges (`🟢 LIVE EPS FEED` vs `⚠️ ARCHIV-DATEN (API OFFLINE)`) in the frontend header to prevent trading on historical cached data without operational awareness.
* *Status*: **Fully Operational (Production Lock)**.
* **Phase 9.0: Radical UI Re-Layout & Accuracy Overhaul**
* *Features*: Destructured the dual-column layout inside `CryptoDemo.tsx` into centered full-width panels. Added human-readable time conversion for pending countdowns, interactive calibration details toggle, individual horizon accuracy tracking, and an accordion detail toggle showing correctness checkmarks for all 5 models.
* *Status*: **Fully Operational (Production Lock)**.
---
@@ -216,7 +219,16 @@ $$P_{\text{Posterior}} = \frac{\alpha_{\text{prior}} + (P_{\text{ML}} \times w)}
#### 5. Walk-Forward Validation & Multi-Model Ensemble
To prevent look-ahead bias and structural overfitting, the system deploys a Walk-Forward Validation framework on a fixed 365-day rolling window across a fleet of 5 machine learning estimators: Random Forest (RF), XGBoost/Gradient Boosting (GB), ElasticNet Logistic Regression (LR), Support Vector Machines (SVM), and Multi-Layer Perceptrons (MLP).
Predictions are generated across three distinct forecast horizons: \(T+1\), \(T+5\), and \(T+10\). To ensure absolute stationarity, all raw asset prices are stripped from the feature space, utilizing only Log-Returns, Rolling Volatility, RSI, Distance to Moving Averages, and Daily Spreads.
Predictions are generated across three distinct forecast horizons: \(T+1\), \(T+5\), and \(T+10\). To ensure absolute stationarity, all raw asset prices are stripped from the feature space. The feature matrix contains asset-specific preprocessed attributes (Log-Returns, Rolling Volatility, RSI, Distance to EMA/SMA, High-Low Spreads) augmented by four high-density intermarket and behavioral control variables:
1. **US Equity Risk Premium Proxy**: Nasdaq Composite Index (`^IXIC`) 1-day log returns.
2. **Safe Haven Real Yield Proxy**: Gold Spot Futures (`GC=F`) 1-day log returns.
3. **Systematic Market Fear Control**: Volatility Index (`^VIX`) raw closing level.
4. **Behavioral Retail Euphoria Matrix**: Crypto Fear & Greed Index (Alternative.me REST API, normalized mapping 0-100).
All intermarket variables are structured and aligned using forward-fill (`ffill()`) and backward-fill (`bfill()`) to resolve weekend stock/futures data gaps.
##### Feature Importance Routing Gateway:
To protect non-linear models (SVM and MLP) from catastrophic overfitting due to noise, the pipeline runs a structural feature selection gateway. Data is passed through a Random Forest selector (`SelectFromModel` with `threshold="mean"`) to extract top-performing features before training the SVM and MLP estimators.
##### Leakage Safeguards (Horizon Cutoff):
For a training window ending at index \(T-1\) and forecasting horizon \(H \in \{1, 5, 10\}\):
@@ -227,7 +239,7 @@ For a training window ending at index \(T-1\) and forecasting horizon \(H \in \{
##### Multi-Tracker Online Learning:
The cockpit maintains 15 independent Beta-Posterior trackers (5 models \(\times\) 3 horizons) persisted inside the client browser. Each tracker is initialized with historical priors and updated dynamically in the background. The expected accuracy is calculated as:
\[\mathbb{E}[\theta] = \frac{\alpha}{\alpha + \beta}\]
where \(\alpha\) represents successes and \(\beta\) represents false alarms, calculated independently for each estimator-horizon pair.
where \(\alpha\) represents successes and \(\beta\) represents false alarms, calculated independently for each estimator-horizon pair. Evaluated performance statistics are compiled on the dashboard in the **Global Performance Metrics Panel**, detailing Horizon Efficiency (Section A) and Estimator Hit Distribution (Section B).
---