AI and Exchange Rate Predictability

Battle Count: 72

Highly relevant to quantitative trading. The paper directly constructs and evaluates a cross-sectional FX trading strategy (AIFX strategy) with Sharpe ratios exceeding 0.7, demonstrates alpha after controlling for standard currency factors (Dollar, Carry, Momentum, Value), and provides a novel signal construction methodology using AI. The look-ahead bias analysis is particularly valuable for practitioners concerned about data leakage. The time-series strategy variant and weighted index specification offer additional implementation options. The finding that predictability is driven primarily by positive news (Strength ratio) has direct portfolio construction implications.

Moderate complexity. The core methodology involves: (1) collecting economic calendar data from Investing.com, (2) constructing structured prompts for LLM API calls, (3) parsing AI outputs into directional signals, (4) computing the AIFX index over lookback windows, (5) constructing cross-sectional portfolios (long top-2, short bottom-2), and (6) evaluating performance against benchmark factors. The main challenges are: API cost management for 174,820 data points, ensuring consistent prompt engineering, handling API rate limits, and replicating the look-ahead bias tests. No code is provided, requiring implementation from scratch.

The paper provides detailed prompts (Prompts 1-8), clear variable construction formulas, and specifies data sources (Investing.com economic calendar, Bloomberg exchange rates). However, reproducibility is limited by: (1) reliance on proprietary API access to GPT-4o and DeepSeek-V3, (2) commercial data subscriptions (Investing.com, Bloomberg), (3) no code repository provided, (4) AI model outputs may vary across API calls and model versions. The methodology is well-documented but full replication requires paid data and API access.