Dynamic Causal Inference in Finance using Meta-RL

1. Project Overview

This project implements a novel data science approach combining Causal Inference and Meta-Reinforcement Learning (Meta-RL) to analyze highly non-stationary time series data.

The core goal is to move beyond simple prediction and achieve dynamic causal inference: measuring how the causal influence between market factors changes over time due to shifting market regimes (non-stationarity).

Component	Role	Novelty
Causal Model	Vector Autoregression (VAR(2))	Used to calculate the time-lagged causal scores (Granger Causality).
Learning Algorithm	Model-Agnostic Meta-Learning (MAML)	Meta-learns the optimal starting point ($\Theta^*$) for the VAR(2) model, enabling rapid adaptation to new causal structures in minutes, not hours.
Key Output	Time-Varying Causal Coefficients	Provides empirical evidence of how market rules evolve across different historical periods.

---

2. Dataset and Variables

We use real-world daily financial data for Apple Inc. (AAPL) spanning over four decades (1980–2022). The raw data is transformed into three key time series variables ($\mathbf{Y}_t$) to establish a robust causal system ($k=3$).

Variable	Symbol	Source/Calculation	Interpretation
Returns	$Y$ (R)	Log Return of Adjusted Close	Outcome: The percentage price change we seek to explain.
Volume	$X_1$ (V)	Log of Trading Volume	Cause 1: Measures market activity/liquidity.
Volatility	$X_2$ (Vol)	Log(High/Low) Range	Cause 2: Measures market risk/uncertainty.

---

3. Project Approach: VAR(2) and MAML

A. The Challenge: Non-Stationarity

Financial markets are non-stationary; the effect of volume on price during a pre-internet market (1980s) is fundamentally different from a high-frequency trading market (2020s). A traditional model would fail by averaging these effects.

B. The Solution: MAML Task Structure

The entire 40-year time series is segmented into sequential Tasks ($T_1, T_2, \dots$). Each task is treated as a separate, unique causal environment.

1. Inner Loop (Adaptation): The VAR(2) model quickly adjusts its causal coefficients ($\phi$) using a small Support Set of data from the current task, reflecting the current market rule.
2. Outer Loop (Meta-Learning): The MAML algorithm optimizes the model's initialization ($\Theta^*$) across all tasks, ensuring the model is always initialized to the most adaptable starting point for any future market shift.

---

4. Causal Coefficients and Inference

The core finding is contained in the Causal Coefficient Matrix ($\mathbf{\Phi}_1$), which is the output of the trained model.

A. Initial Causal Matrix ($\Theta^*$)

This matrix represents the Meta-Learned Master Strategy—the average week-to-week influence one variable has on another across the entire dataset.

Coefficient	Value (Example)	Interpretation (Granger Causality)
$\mathbf{\Phi}_{1}[\text{R, V}]$	+0.377	STRONG POSITIVE CAUSAL SCORE: On average, higher trading volume last week significantly contributes to higher returns this week.
$\mathbf{\Phi}_{1}[\text{R, R}]$	-0.228	NEGATIVE EFFECT: Past price increases lead to slight downward pressure (mean-reversion) this week.

B. Dynamic Inference (The Novelty)

The causal_track_df.csv file contains the dynamic coefficients. By plotting these, we show the causal score for the $\text{Volume} \to \text{Returns}$ link is not constant across all time, but fluctuates (e.g., from $+1.0$ to $-0.1$) proving that the market's fundamental rules are unstable and adaptive.

---

5. Running the Project

A. Setup

1. Create Environment: Ensure you have Python and PyCharm/VS Code.
2. Install Dependencies: Run the following command using the requirements.txt file:

   pip install -r requirements.txt

B. Execution

1. Data Placement: Place your AAPL.csv file in the main project directory.
2. Training: Run the training script to optimize the MAML model and generate results. A fixed seed is used for reproducibility.

   python maml_trainer.py

3. Visualization: Launch the Streamlit dashboard to view the dynamic results.

   streamlit run app_dashboard.py

---

6. Final Results and Applications

The project successfully delivered a working MAML-VAR(2) framework that achieves stable meta-learning convergence on non-stationary financial data.

Final Deliverables

• theta_star.pt: Saved optimized initial parameters.
• causal_track_df.csv: Time series data of the evolving causal coefficients.
• Streamlit Dashboard: Interactive visualization of the dynamic causal scores, proving the existence of non-stationary causal regimes.

Other Applications

The Meta-RL Causal Inference framework developed here is highly generalizable and can be applied to other complex, evolving systems:

• Public Health: Tracking the time-varying effect of vaccination rates on disease transmission as the virus mutates.
• Climate Science: Inferring the evolving causal relationship between greenhouse gas concentrations and extreme weather events.
• Policy Economics: Measuring how the effectiveness of a fiscal stimulus policy changes as consumers and markets adapt over time.