PnL Attribution & Alpha Decay Framework

Strategy forensics for the FinBERT sentiment signal on AAPL.
Decomposes where a signal's edge comes from and how fast it decays.

Quick start

git clone https://github.com/maanitmehta/pnl-attribution.git
cd pnl-attribution
python3 -m venv venv
source venv/bin/activate      # Windows: venv\Scripts\activate
pip install -r requirements.txt
python app.py

Open http://localhost:8050 in your browser.

---

What it does

The dashboard loads a trade log for AAPL — either synthetic (2022–2024, 200 signals with engineered IC decay) or real FinBERT signals from the sentiment_signals SQLite database (2025–2026, 29 signals). Select the dataset with the radio buttons at the top.

Panel 1 — Strategy Overview

Simulates a portfolio that acts on each FinBERT signal at a 1-day holding period (full capital re-investment, no leverage). Compares against a buy-and-hold baseline.

Metric	Plain-English meaning
Sharpe Ratio	Risk-adjusted return: higher is better. Annualised as mean(trade PnL) / std(trade PnL) × √252.
Hit Rate	Fraction of active trades where the signal predicted the correct direction.
Max Drawdown	Worst peak-to-trough loss in the equity curve — a measure of tail risk.
Avg PnL / Trade	Mean return per trade in basis points (1 bps = 0.01%).

Panel 2 — PnL Attribution Waterfall

Breaks total PnL into three components:

Component	What it measures
Gross Signal Alpha	What the signal would have earned if you could execute at the signal-day close price — the pure predictive edge.
Timing Slippage	The gap between next-day open (actual fill) and signal-day close (theoretical fill). Overnight moves and bid-ask crossings create this cost.
Execution Drag	Fixed cost of 7 bps per trade (5 bps bid-ask spread + 2 bps market impact).
Net PnL	Gross Alpha + Timing Slippage + Execution Drag.

The scatter plot shows raw signal score vs forward return at the chosen horizon, with an OLS trend line (slope ≈ IC × volatility scaling).

Panel 3 — Alpha Decay Curve (the key chart)

Plots the Information Coefficient (IC) at each holding horizon: 1, 2, 3, 5, 10, 20 days.

IC = Spearman rank-correlation between signal_score and forward return.

• IC = 1.0 → perfect ranking (every strong-positive score is a winner)
• IC = 0.0 → no predictive power
• IC = −1.0 → perfectly contrarian

We use Spearman (rank-based) rather than Pearson because equity returns are fat-tailed and rank ordering is what a portfolio manager actually cares about.

The 95% confidence bands use the asymptotic standard error √((1−IC²)/(N−2)). The red dashed line marks the first horizon where IC drops below statistical significance (p > 0.05) — this is the maximum useful rebalancing frequency for this signal.

Green nodes = statistically significant. Red nodes = not significant.

Panel 4 — Rolling Signal Quality

Rolling IC computed over a sliding window of trades (default: 60).

• Green shading = IC > 0 (signal is adding value in this window)
• Red shading = IC < 0 (signal is actively harmful — regime change alert)
• Amber dashed line = rolling hit rate (50% = coin flip)

Use the window slider to control sensitivity. A narrow window (20 trades) shows short-term IC volatility; a wide window (100 trades) reveals structural decay.

---

Project layout

pnl_attribution/
├── app.py             Dash app — layout + callbacks
├── data_pipeline.py   Data loading, IC computation, PnL decomposition
├── cache/             yfinance price CSVs (auto-created, gitignored)
└── README.md

Key functions in data_pipeline.py

Function	What it does
fetch_prices()	yfinance OHLCV download, cached to cache/
load_real_signals()	Reads FinBERT signals from sentiment_signals/data/sentiment.db
generate_synthetic_trades()	Generates signals with target IC ≈ 0.18 at 1d, natural √-decay at longer horizons
build_trade_log()	Main entry: tries real, falls back to synthetic
compute_ic_table()	IC + t-stat + p-value at each horizon
compute_rolling_ic()	Rolling 60-trade IC for Panel 4
compute_pnl_attribution()	Gross / slippage / drag waterfall decomposition

---

How the synthetic IC decay works

The synthetic signal is constructed in rank-normal space:

z_ret  = Φ⁻¹( rank(ret_1d) / (N+1) )          # rank-normalise 1d return
signal = ρ · z_ret  +  √(1−ρ²) · ε             # blend with noise at target IC ρ
score  = tanh(signal / 2)                        # squash to (−1, +1)

This gives IC ≈ ρ at the 1-day horizon. At longer horizons, because multi-day returns accumulate independent daily noise, IC decays as IC_h ≈ IC_1 / √h — a classical result in quantitative finance. The decay crosses statistical significance at around 3–5 days for a realistic IC of 0.18, which is the critical insight the Keyrock alpha decay chart captures.

---

Built as a quant research portfolio piece — demonstrates PnL attribution, IC analysis, and signal forensics methodology used in systematic trading.