Unlocking Alpha in G7 Currency Markets with Financial Times and RavenPack Core News Sentiment Ensembling

The empirical evidence is consistent: weak and highly-orthogonal signals, when combined from independent sources and filtered for agreement, generate more reliable alpha than any single source alone.

News sentiment ensembling offers a rigorous framework for translating macroeconomic news into persistent currency momentum. Our latest research demonstrates how combining country-level macroeconomic news sentiment from RavenPack Core News and the Financial Times, through a cross-validated ensemble approach, can significantly enhance risk-adjusted performance across the G7 FX futures markets.

Since 2004, this multi-source approach has improved Information Ratios from 0.48 to 0.81 while maintaining a portfolio turnover of approximately eight trading days, adding 108 basis points of annualized excess return over a standalone RavenPack Core News strategy.

The challenge: weak signals and single-source dependence

Macroeconomic news sentiment signals exhibit low and unstable information coefficients, making them difficult to implement in currency rotation strategies without introducing excessive turnover or unwanted noise. Single-source approaches are particularly vulnerable: any given news provider covers events selectively, creating blind spots that weaken signal reliability and inflate the risk of false conviction.

Traditional currency trend-following strategies that rely on price momentum alone leave a significant information gap. News sentiment is structurally uncorrelated with price momentum, a characteristic that creates a genuine diversification opportunity, but only if that sentiment can be extracted reliably.

The solution: cross-source sentiment ensembling

The solution lies in combining two independent, complementary news sources and filtering signals based on cross-sectional agreement between them. The ensemble process validates consensus across sources, reinforcing extreme signals and minimizing noise using a non-linear softmax filter.

Our approach: sentiment momentum with a voting mechanism

Our framework implements signal construction and combination at two distinct stages:

1. Individual signal construction: For each source separately, we aggregate the Composite Sentiment Score (CSS) daily at the country level across the 27 countries and regions corresponding to G7 currency pairs. Scores are exponentially smoothed over a six-month lookback window, weighting recent news more heavily, and currencies are ranked daily from highest to lowest macroeconomic sentiment.

2. Ensemble voting mechanism: The two ranked signals are combined using a hyperbolic tangent function: a non-linear, softmax-like approach validates cross-sectional agreement across sources, reinforcing extreme signals and reducing noise and low-conviction signals.

Signal construction

Sentiment scores are sourced from RavenPack Edge Analytics and RavenPack FT Analytics, covering macroeconomic topics including GDP, employment, CPI, interest rates, and consumption across all G7 developed market currencies. Non-neutral events are retained, machine-generated content is excluded, and only highly reputable sources are included. The result is a clean, daily currency ranking that feeds directly into a dollar-neutral, long-short FX futures strategy daily rebalanced at the close of the London trading session.

Performance results

Testing across multiple portfolio sizes, time decay lengths, and the full 2004–2025 sample period reveals:

• Information Ratio: 0.48 → 0.81
• Annualized Excess Return: +108 bps over RavenPack Core News alone
• Average Return Spread: ~260 bps over a bi-weekly effective holding period

The ensemble consistently delivers superior risk-adjusted returns across the full parameter space. Even at smaller portfolio concentrations, the strategy maintains attractive Information Ratios while keeping effective holding periods stable.

Why ensembling works. The mechanics

The effectiveness of cross-source sentiment combination stems from several reinforcing principles:

• Noise reduction through independence. RavenPack Core News and the Financial Times cover macroeconomic events through distinct editorial processes, source bases, and publication cadences. Their average rank correlation is only 10–14%, confirming that they carry genuinely different information. Combining them filters idiosyncratic noise while preserving the shared signal.

• Statistical confirmation as a return predictor. The key finding is not simply that the ensemble outperforms. It is those days with higher Spearman rank correlation between the two signals are statistically linked to higher next-day portfolio returns (p-value = 0.00). Signal agreement is itself a predictive variable, not just a construction choice.

• Non-linear amplification. The hyperbolic tangent function applied to the combined signal emphasizes extreme, high-conviction rankings while dampening ambiguous mid-range positions. This mirrors how professional discretionary traders weigh consensus: conviction compounds, uncertainty recedes.

• Turnover stability. Because the ensemble is built on smoothed sentiment momentum rather than raw daily scores, portfolio turnover remains controlled at roughly eight trading days — comparable to a single-source strategy, but with materially better risk-adjusted returns.

Download the full research

Ensemble news sentiment construction demonstrates that well-structured multi-source combination can unlock meaningful alpha from macroeconomic news data while maintaining institutional-grade turnover discipline. The improvement from an Information Ratio of 0.48 to 0.81, achieved without additional complexity in portfolio construction, highlights a compelling risk-return case for active FX managers and systematic currency overlay strategies.

For full methodology, signal decomposition, filter sensitivity analysis, and comprehensive backtesting results across all G7 currency pairs, download the complete white paper.