PaperlayerView exact revision on arXiv

Source-linked AI summary

Forecasting Intraday USD/CAD Exchange Rate with News-Derived Monetary-Policy Signals

Maya Kodeih, Aliaa Alnaggar, Mucahit Cevik

arXiv:2609.32773v1cs.AIcs.LGq-fin.CP

TL;DR

The paper asks which dimensions of monetary-policy communication contribute forecasting information beyond aggregate accuracy comparisons. It structures monetary-policy news into signals and evaluates them with rolling-window forecasting and statistically controlled attribution. The results concentrate predictive value in timing and targeted activity, with complementary information from LLM-derived sentiment.

  • Problem

    Prior work mainly evaluates textual information through aggregate forecasting accuracy, with limited evidence on which monetary-policy communication dimensions contribute to improvements.

    Source: §I · p. 1, §II · p. 2

  • Method

    The paper transforms central-bank news into communication signals and evaluates them through rolling-window forecasting, feature- and group-level attribution, and false-discovery-rate-controlled testing.

    Source: §I · p. 1, G. Feature Attribution Analysis · p. 5

  • Results

    Removing is_news_hour reduces out-of-sample explanatory power by approximately 1.2 percentage points; attribution also finds positive activity contributions and complementary group-level sentiment information.

    Source: D. Feature-Level Attribution Results · p. 8, E. Group-Level Attribution Results · p. 9

  • Takeaways & Limitations

    Within the tested framework, communication-based forecasting value extends beyond sentiment alone, making channel attribution informative alongside aggregate accuracy.

    Source: §V · p. 10

    The evaluation is restricted to USD/CAD and Federal Reserve and Bank of Canada communications, limiting direct generalization to other currencies and monetary-policy regimes.

    Source: §V · p. 10

Abstract

from arXiv · show

Monetary-policy announcements and central-bank communications play a central role in foreign exchange markets, yet their qualitative, unstructured form makes their forecasting value difficult to quantify. While prior research has largely focused on sentiment extracted from financial news, comparatively little is known about the relative contribution of different dimensions of monetary-policy communication. Existing studies primarily evaluate whether textual information improves overall forecasting performance but provide limited insight into which communication channels drive such improvements. To address this gap, this paper introduces a statistical attribution methodology that decomposes monetary-policy communication into interpretable channels and quantifies their incremental forecasting contribution under false-discovery-rate control. Monetary-policy news is transformed into structured communication signals using large language models (LLMs) and temporal feature engineering. These signals are evaluated using rolling-window experiments with tree-based machine-learning models. The results show that monetary-policy communication contains measurable predictive information. Attribution analysis shows that predictive value is concentrated in a small subset of signals, with communication timing providing the strongest individual feature-level contribution, targeted communication-activity measures also contributing positively, and LLM-derived sentiment providing complementary information at the group level. The findings indicate that communication-based forecasting value extends beyond sentiment alone and that attribution, rather than aggregate accuracy alone, is central to evaluating news-derived signals.

I. INTRODUCTION

Prior FX-news studies largely evaluate aggregate forecasting accuracy, leaving limited evidence about which monetary-policy communication dimensions drive improvements. This paper introduces statistically controlled attribution and finds predictive value concentrated in a small subset of signals.

II. LITERATURE REVIEW

  • Prior studies primarily assess textual information through aggregate accuracy, while evidence about which communication dimensions contribute to forecasting improvements remains limited.The attribution gap includes sentiment, timing, communication activity, policy stance, and persistence.Source: §I · p. 1, §II · p. 2

I. INTRODUCTION

  • The framework combines LLM-based sentiment extraction, temporal feature engineering, rolling-window forecasting, attribution, and false-discovery-rate-controlled testing.Its signals cover sentiment, activity, timing, policy stance, and information persistence.Source: §I · pp. 1–2
  • Communication timing provides the strongest individual feature-level contribution, targeted communication-activity and non-neutral policy signals also contribute positively, and LLM-derived sentiment adds group-level information.The paper reports that predictive value is concentrated in a small subset of engineered features.Source: §I · p. 2

II. LITERATURE REVIEW

  • The study extends prior USD/CAD work on daily and weekly macro-financial variables to intraday forecasting with news- and LLM-derived central-bank signals.It also evaluates directional and probabilistic forecasts alongside point forecasts.Source: §II · p. 2

III. METHODOLOGY

The methodology evaluates communication signals for intraday USD/CAD forecasting through a staged attribution workflow. Its dataset aligns filtered monetary-policy news with hourly exchange-rate observations for rolling-window evaluation.

III. METHODOLOGY

  • The integrated workflow targets channel-level attribution through data preparation, signal extraction and engineering, forecasting, attribution, and statistical testing.It includes deterministic benchmark evaluation and probabilistic forecasting with uncertainty estimation.Source: §III · p. 2

A. Data Collection and Processing

  • News articles covering macroeconomic developments and central-bank communications are filtered for relevance to monetary-policy expectations.The filtering is designed to retain articles relevant to monetary-policy information.Source: A. Data Collection and Processing · p. 2
  • The sample contains 927 monetary-policy-related articles aligned with 17,517 hourly USD/CAD observations from 2024-06-13 to 2026-06-12.Each rolling window uses an 85/15 train-test split, with approximately 4,324 training and 764 out-of-sample observations.Source: A. Data Collection and Processing · pp. 2–3
  • Articles are filtered using a monetary-policy vocabulary, deduplicated, timestamped, and aggregated hourly to align news features with intraday exchange rates.Exchange-rate observations during weekend closures are forward-filled using the most recent available rate.Source: A. Data Collection and Processing · p. 3

B. Monetary-Policy Sentiment Extraction

The paper uses an LLM to classify central-bank articles by monetary-policy stance, then converts article-level labels into hourly signals. Feature engineering adds communication timing, activity, persistence, sentiment differentials, and market variables.

B. Monetary-Policy Sentiment Extraction

  • GPT-4o-mini classifies each article’s implied monetary-policy stance as hawkish, dovish, neutral, or mixed, separately for Federal Reserve and Bank of Canada communications.A manual review of 100 articles found labels generally consistent with the expressed stance in that validation sample.Source: B. Monetary-Policy Sentiment Extraction · p. 3

C. Feature Engineering

  • The engineered feature set combines sentiment-derived variables, communication timing and activity, persistence indicators, and conventional market variables.Examples include rolling communication counts, news-arrival indicators, sentiment differentials, impulse and decay features, USD/CAD lags, volatility, interest-rate spreads, macroeconomic momentum, and WTI prices.Source: C. Feature Engineering · p. 3
  • A central signal is the Federal Reserve–Bank of Canada monetary-policy sentiment differential, with impulse variables representing immediate information-arrival effects.The article-level labels are also aggregated hourly into sentiment differentials, non-neutral communication indicators, persistence measures, and rolling summaries.Source: C. Feature Engineering · pp. 3–4, B. Monetary-Policy Sentiment Extraction · p. 3
  • Exponentially decaying sentiment variables capture information persistence, while rolling sentiment indicators use multiple lookback windows.The decay rate is controlled by λ, and k denotes the rolling-window length.Source: C. Feature Engineering · p. 4

D. Forecasting Models

The paper uses tree-based ensembles for USD/CAD forecasting, fitting separate direct models by horizon and evaluating forecasts against moving-average and random-walk benchmarks. It also tests directional predictability with classification models and benchmarks designed to account for class imbalance and persistence.

D. Forecasting Models

  • Extra Trees, XGBoost, and LightGBM model nonlinear relationships and feature interactions; separate direct regressors for each horizon avoid recursive error propagation.The same model families are also used for directional classification, and quantile regression extends deterministic forecasts for probabilistic evaluation.Source: D. Forecasting Models · p. 4
  • The no-change random walk is the principal level-forecast benchmark, alongside a moving-average predictor based on the three latest observations in each rolling training window.The paper frames its main objective as attribution rather than benchmark dominance.Source: D. Forecasting Models · p. 4

E. Directional Forecast Evaluation

  • Directional classification estimates the probability of an upward USD/CAD movement and converts it to a binary prediction using a decision threshold.The directional target distinguishes upward movement from the alternative outcome.Source: E. Directional Forecast Evaluation · p. 4
  • Directional evaluation uses accuracy, balanced accuracy, macro F1, and AUC, comparing predictions with majority-class and persistence baselines.Balanced accuracy and benchmark-relative performance help distinguish predictive signal from class imbalance or simple persistence.Source: E. Directional Forecast Evaluation · p. 4

F. Probabilistic Forecast Evaluation

Probabilistic forecasting estimates conditional prediction intervals, while feature and group ablations assess the incremental information in communication signals. The ablation procedure retrains models after removals and measures performance degradation relative to the preceding step.

F. Probabilistic Forecast Evaluation

  • Conditional prediction intervals extend single-value forecasts to characterize predictive uncertainty.LightGBM and XGBoost estimate conditional quantiles directly, while Extra Trees derives empirical quantiles from ensemble-member predictions.Source: F. Probabilistic Forecast Evaluation · p. 4
  • Uncalibrated quantile forecasts are the primary probabilistic comparison, assessed with pinball loss, weighted interval score, and empirical interval coverage.Conformal calibration is evaluated separately using held-out calibration residuals and is excluded from the feature-specification comparison in Table III.Source: F. Probabilistic Forecast Evaluation · p. 4

G. Feature Attribution Analysis

  • Sequential feature ablation measures out-of-sample performance change after removing predictors and retraining under the same rolling-window design.This estimates incremental value under the full pipeline, unlike model-specific importance, and allows interactions among remaining features to adjust after retraining.Source: G. Feature Attribution Analysis · pp. 4–5
  • Positive degradation after removal indicates useful predictive information, negligible change suggests redundancy, and negative degradation indicates redundant or detrimental information.Feature-level ablation measures individual variables; group-level ablation jointly removes related variables to assess collective contribution.Source: G. Feature Attribution Analysis · p. 5

H. Statistical Significance Testing

Ablation effects are tested with bootstrap confidence intervals and Benjamini–Hochberg false-discovery-rate correction. A feature or group is supported only when its interval excludes zero and its adjusted q-value meets the predefined threshold.

H. Statistical Significance Testing

  • Bootstrap confidence intervals estimate the sampling distribution of feature- and group-level attribution metrics from resampled rolling-window performance differences.The tests use nonparametric bootstrap procedures.Source: H. Statistical Significance Testing · p. 5
  • Benjamini–Hochberg correction is applied across feature-level two-sided p-values computed from pooled stepwise ∆R2 observations across evaluated model-horizon combinations.The procedure controls the expected proportion of false discoveries while offering greater power than family-wise error-rate corrections.Source: H. Statistical Significance Testing · p. 5
  • An attribution effect is statistically supported only when its bootstrap confidence interval excludes zero and its FDR-adjusted q-value meets the predefined significance threshold.Applying both criteria makes reported effects more conservative and reduces the likelihood of false-positive discoveries.Source: H. Statistical Significance Testing · p. 5

I. Experimental Design

The experiments use rolling-window forecasts across deterministic, probabilistic, and directional tasks and horizons of 1, 3, 8, and 24 hours. In level forecasting, machine-learning models do not outperform the no-change random walk, although their relative performance gap narrows at longer horizons.

I. Experimental Design

  • Rolling-window evaluation trains on historical observations and tests on subsequent data, so each prediction uses only information available at forecast time.The design avoids look-ahead bias and applies consistent preprocessing, feature engineering, benchmark comparisons, and statistical testing across horizons and configurations.Source: §I · pp. 5–6
  • The three tasks are deterministic level prediction, probabilistic prediction intervals, and directional binary classification; attribution uses deterministic point forecasts.Probabilistic forecasting separately checks whether the communication hierarchy found through deterministic attribution persists under uncertainty-aware forecasting.Source: §I · p. 5
  • The evaluation covers 1-, 3-, 8-, and 24-hour horizons, using RMSE, MAPE, and R2 for deterministic forecasts and task-specific probabilistic and classification metrics.The directional task also compares against majority-class and persistence benchmarks.Source: §I · p. 5

A. Forecasting Performance Across Horizons

  • None of the evaluated machine-learning models outperforms the no-change random walk on level-based metrics; the relative gap narrows at longer horizons.Across horizons, machine-learning NRMSE is higher and R2 lower than the corresponding random-walk results; reported model R2 ranges from 0.844–0.915.Source: A. Forecasting Performance Across… · p. 6, Table II · p. 6
  • Tree-model RMSE generally decreases after early ablations remove weak or redundant sentiment and directional features, especially at 1- and 3-hour horizons.At longer horizons, Extra Trees remains comparatively stable while boosting models become more sensitive to cumulative removal, indicating horizon- and interaction-dependent responses.Source: A. Forecasting Performance Across… · p. 6

B. Probabilistic Forecasting Results

The Step 7 communication-aware specification improves probabilistic scoring over the full feature set at every horizon, but empirical interval coverage remains a qualification, especially at longer horizons.

B. Probabilistic Forecasting Results

  • Uncalibrated intervals under-cover at every horizon: nominal 50% coverage ranges from 30.85% to 44.01%, and nominal 90% coverage from 53.72% to 63.23%.Conformal calibration is evaluated separately and is not reflected in Table III's coverage values.Source: B. Probabilistic Forecasting Results · p. 6
  • Pinball loss falls by approximately 55.6%, 46.7%, 29.3%, and 10.8% at the 1-, 3-, 8-, and 24-hour horizons, respectively, under Step 7 versus the full feature set.The comparison uses uncalibrated quantile forecasts; weighted interval score exhibits a similar pattern.Source: Table III · p. 6, B. Probabilistic Forecasting Results · p. 6
  • Step 7 does not uniformly improve nominal 90% coverage: it rises from 53.72% to 68.92% at 1 hour but falls at longer horizons.The models score better but become less conservative as the forecasting horizon increases.Source: B. Probabilistic Forecasting Results · pp. 6–7
  • Figure 2 tracks mean RMSE across cumulative ablation steps at 1-, 3-, 8-, and 24-hour horizons, with each step removing one feature from the full set.Lower RMSE indicates better forecasting performance.Source: Figure 2 · p. 7
  • At 1 hour, the median forecast closely follows USD/CAD and intervals capture much short-term variation; at 24 hours, rapid moves are less well tracked and fall outside intervals more often.This horizon contrast is consistent with the reduced empirical coverage reported in Table III.Source: B. Probabilistic Forecasting Results · p. 7

C. Directional Forecasting Performance

Directional models outperform both benchmark strategies across all horizons, with benchmark-relative gains increasing at longer horizons. Feature-level attribution shows that timing and targeted communication activity contribute most reliably, while broad volume and some long-horizon recency measures are redundant.

C. Directional Forecasting Performance

  • 63.22% to 66.55% directional accuracy across horizons exceeds both benchmarks, with majority-class gains rising from approximately 3.0 to 12.1 percentage points.The gains are approximately 3.0, 3.6, 8.7, and 12.1 percentage points at the 1-, 3-, 8-, and 24-hour horizons, respectively.Source: C. Directional Forecasting Performance · p. 7
  • Directional accuracy changes most during early cumulative ablation steps, then remains comparatively stable; effects vary by model and horizon.The sequence measures cumulative feature removals, so its effects are not fixed measures of feature importance.Source: C. Directional Forecasting Performance · p. 7, Figure 4 · p. 8
  • Balanced accuracy of 0.626–0.721, macro F1 of 0.604–0.660, and AUC of 0.634–0.744 indicate discrimination beyond class imbalance.The strongest balanced-accuracy performance occurs at 1 hour, while meaningful class discrimination remains at 24 hours.Source: C. Directional Forecasting Performance · p. 7

D. Feature-Level Attribution Results

  • Removing is_news_hour reduces out-of-sample explanatory power by approximately 1.2 percentage points, the strongest supported individual feature contribution.The change is modest relative to evaluated-model R2 values above 0.86 across horizons.Source: D. Feature-Level Attribution Results · p. 8
  • Recent non-neutral Federal Reserve communication, Federal Reserve article counts, and Bank of Canada communication activity contribute positively after global FDR correction.These activity measures are reported as more reliable forecasting information than broad sentiment measures alone.Source: D. Feature-Level Attribution Results · p. 8
  • Aggregate article counts and long-horizon recency measures are often redundant once targeted central-bank communication and news-arrival indicators are included.The news_articles_24h and news_articles_6h variables have negative ∆R2; hours_since_boc_news_168h and hours_since_news_168h appear to duplicate other indicators.Source: D. Feature-Level Attribution Results · p. 8

E. Group-Level Attribution Results

Group-level attribution finds positive average contributions from communication timing and LLM-derived sentiment, while broad news volume contributes negatively. Together, the group and feature results indicate that timing and targeted communication activity carry useful information, with sentiment adding complementary information.

E. Group-Level Attribution Results

  • Communication timing has a positive average contribution, consistent with is_news_hour being the strongest individual contributor.The paper identifies timing of policy-relevant information as one of the most informative short-horizon signals.Source: E. Group-Level Attribution Results · p. 9
  • The news-volume group has a negative average contribution, consistent with aggregate article counts becoming redundant alongside targeted central-bank indicators.In these experiments, source and timing measures provide more useful information than broad overall news-volume measures.Source: E. Group-Level Attribution Results · p. 9
  • LLM-derived sentiment has a positive average group contribution, providing complementary predictive information beyond timing and activity alone.Individual sentiment variables are generally modest, though several representative measures have positive contributions.Source: E. Group-Level Attribution Results · p. 9
  • The LLM pipeline constructs representations of policy confidence, directional implications, non-neutral communication, persistence, and sentiment dynamics.Some activity measures, especially non-neutral communication counts, depend on information extracted during language-model interpretation.Source: E. Group-Level Attribution Results · p. 9
  • Model importance and cumulative ablation can rank features differently: oil price, rate spread, and sentiment-decay features rank high in importance but add little under ablation.is_news_hour has low model importance but the largest ablation effect; the paper distinguishes model utilization from incremental forecasting value.Source: E. Group-Level Attribution Results · p. 9

V. CONCLUSION AND DISCUSSION

The paper concludes that intraday USD/CAD predictive value is concentrated in a few monetary-policy communication signals, and that cumulative ablation provides a different perspective from conventional feature importance. Its findings are bounded by the tested currency, central banks, attribution sequence, forecasting objective, and LLM choices.

V. CONCLUSION AND DISCUSSION

  • Predictive value concentrates in a small subset of signals: timing is the strongest individual contributor, targeted central-bank activity contributes positively, and group-level sentiment adds complementary information.Broad news-volume measures are largely redundant.Source: §V · p. 10
  • Cumulative ablation measures incremental forecasting value, unlike feature-importance measures, which describe how models use available information.The reduced feature set improves probabilistic scores, but sharpness and interval calibration must be judged separately.Source: §V · p. 10
  • The evaluation covers USD/CAD and Federal Reserve and Bank of Canada communications, limiting direct generalization to other currencies and policy regimes.It evaluates statistical forecasting rather than economic value, excluding transaction costs, market frictions, and trading profitability; LLM features also depend on prompt design, model selection, and classification uncertainty.Source: §V · p. 10
  • Future work can test the stability of the communication hierarchy across more currency pairs, central banks, and multilingual sources.The paper also suggests alternative attribution comparisons, formal benchmark tests, realistic trading evaluations, and more advanced probabilistic calibration.Source: §V · p. 10

Related papers

Loading 2609.32773v1…