D’s Method — Volatility-Scaled Sentiment Fusion for Intraday Trading (MFT)

D’s Method is a research-grade intraday trading system designed for medium-frequency trading (MFT) with a 10-minute prediction horizon.

The core idea is simple but powerful:

Numeric models predict structure. Sentiment does NOT predict returns directly — it modulates the numeric signal based on volatility and regime.

This repository contains a fully working, end-to-end implementation:

• strict no-leakage design
• volatility-aware fusion
• sentiment decay and confidence handling
• backtesting, live simulation, and dashboard support

---

1. Core Equation (D’s Method)

The final return prediction is:

Interpretation:

• f_num(X_t) → numeric base prediction from price/volume features
• σ_t → realized short-horizon volatility
• S_c, S_m, S_i → company, market, and index sentiment
• w_k → dynamic sentiment weights
• θ · S_i · S_c → cross-interaction (alignment vs conflict)

Sentiment influence automatically shrinks in low volatility and amplifies in high volatility — no hard switches.

---

2. High-Level System Pipeline

This diagram shows:

• independent numeric and sentiment pipelines
• late-stage fusion (post-prediction)
• execution after volatility normalization

---

3. Data Flow & Verification Loop

Key properties:

• predictions are stored before outcomes are known
• realized returns at t+10 are used only for verification
• financial metrics and ML metrics are tracked separately

---

4. Deployment & Cloud Architecture

Designed for:

• 24×7 runtime
• market-hours-only execution
• auto-restart and health checks
• cloud-friendly deployment (Render / Railway / Cron)

---

5. Numeric Model (f_num)

Features (14 total)

• Returns: ret_1, ret_3, ret_6
• Price structure: hl_range, clv, ema_gap
• Dynamics: ret_accel
• Volatility: vol_10, vol_30, vol_ratio, vol_norm
• Liquidity: pv
• Seasonality: minute_of_day

Design rules

• FIFO rolling buffer (40 rows in live mode)
• buffer resets at day boundaries
• no future information allowed

The numeric model is intentionally simple and stable (e.g., Ridge / Linear / Tree-based).

---

6. Sentiment System

Sources

• Company news
• Market / sector news
• Index (NASDAQ / QQQ) news

Processing

• RSS ingestion
• NLP model (FinBERT or equivalent)
• confidence-weighted scoring
• exponential time decay

Sentiment is never used raw.

---

7. Execution Logic

Volatility-normalized signal

z_t = r̂_(t+1) / σ_t

Trade rules

• Long if z_t > entry_threshold
• Short if z_t < -entry_threshold
• Exit using hysteresis (lower exit threshold)

Risk controls

• Max 1000 decision points per trading day
• no cross-day state leakage
• fractional position sizing

---

8. Backtesting (Ablation-Safe)

The system supports controlled variants:

Variant	Description
V0	Numeric only
V1	Numeric + linear sentiment
V2	Numeric + interaction term
V3	Full D’s Method

Metrics:

• Directional Accuracy
• Information Coefficient (IC)
• RMSE
• PnL, Sharpe, Drawdown

If sentiment = 0 ⇒ all variants collapse to V0 (expected and verified behavior).

---

9. Live Operation

• Runs as a 24×7 service
• Trades only during market hours (Mon–Fri, 9:30–16:00 NY)
• Numeric data: yfinance (10-min bars)
• Sentiment: RSS feeds
• Default mode: paper trading

This is MFT, not HFT. Designed for robustness, not microsecond latency.

---

10. Project Structure

core/        # fusion, volatility, sentiment, rolling buffer
models/      # numeric model (f_num)
backtest/    # V0–V3 backtests
execution/   # execution & hysteresis
live/        # 24×7 live runner
dashboard/   # Streamlit monitoring
docs/        # diagrams & equation images
data/        # local data (git-ignored)

---

11. Status

✔ End-to-end system complete ✔ No look-ahead bias ✔ Live-ready architecture ✔ Research-grade implementation

---

12. Disclaimer

This project is for research and educational purposes only. It is not financial advice and is not intended for live trading with real capital without extensive independent validation.