Evidence from microstructure and machine learning research
Research that addresses high-frequency foreign exchange markets indicates that alternative data such as tick-level order-book information and real-time news feeds can improve very short-term forecasting accuracy. Richard K. Lyons, University of California, Berkeley, demonstrated through microstructure analysis that order flow carries predictive information about near-term price moves, which traditional end-of-day datasets miss. Marcos López de Prado, Cornell University, argues that combining sophisticated machine learning techniques with richer datasets can extract transient signals that conventional econometric models overlook. These gains are commonly incremental and depend heavily on model design and validation procedures rather than simply adding more data.
Types of alternative data and how they help
Order-book snapshots, executed-trade ticks, news sentiment, search-engine queries, and payment-rail flows are the principal alternative sources used in short-term FX work. Order-flow captures immediate supply and demand imbalances that translate into price pressure. Real-time text analytics on financial news and social media accelerate detection of information shocks. Search-query trends can anticipate retail demand shifts in specific currency corridors. However, the usefulness of each source varies by currency pair, time of day, and market participant mix.
Causes of improvement and practical consequences
The principal cause of improved short-term forecasting is that alternative data records market frictions and information arrival at higher frequency than macroeconomic releases. As algorithmic trading proliferated, the signal contained in microstructure features became more exploitable. Consequences include tighter short-term execution strategies, but also greater susceptibility to feedback loops and liquidity evaporation during stress. Jon Frost, Bank for International Settlements, highlights governance, privacy, and concentration risks when large firms control unique data streams. In emerging-market currencies, coverage gaps and language-specific sentiment extraction make gains more uneven and culturally contingent.
Limitations and best-practice considerations
Empirical improvements documented in the literature are often nonstationary and fragile out of sample. López de Prado emphasizes rigorous out-of-sample testing and addressing overfitting through proper cross-validation. Data quality, latency, and regulatory constraints can erase apparent advantages once transaction costs and slippage are considered. Ultimately, alternative data can improve short-term FX forecasts, but reliable gains require expert feature engineering, institutional-grade data governance, and continuous model assessment to remain valid across changing market regimes.