CoreFX's event-impact analysis was built to answer a specific question: when a scheduled economic release occurs, does it produce a statistically meaningful signal for currency direction — or is the reaction indistinguishable from ordinary market noise. Rather than assume an answer, the system was designed to test it directly, using three years of historical calendar and price data.
The initial design choice — comparing the released figure against its forecast — is the industry-standard approach, and the one this system deliberately moved away from.
Comparing a release's actual figure against its forecast — beat the consensus, currency strengthens; miss it, currency weakens — is how the well-known economic surprise indices are constructed, and is standard practice across much of the industry.
The approach has a structural weakness at the retail data layer, however. The forecast is not a single, objective figure. Each data provider aggregates a different panel of contributing economists, on a different survey schedule, and reports a different consensus for the same release. A forecast sourced from one provider is not necessarily what the broader market was actually pricing in. In addition, a clean beat does not reliably move price in the expected direction, since the market's positioning ahead of the release is frequently already aligned with the outcome.
The system was rebuilt around actual-vs-previous: comparing a release's figure against the prior period's own figure, rather than against a forecast. Both values are hard, released data points — the comparison carries no third-party estimate at any stage.
Initial results were promising: a set of real, interpretable relationships, several consistent with basic economic intuition. On their own, however, these results were not sufficient grounds for trust.
Every relationship identified above had been discovered and measured using the same historical window — a well-documented failure mode in this kind of analysis. A pattern measured and validated on identical data will appear stronger than it is, regardless of whether the underlying relationship is genuine.
The historical window was divided into four equal chronological blocks of approximately ten months each. The first block was reserved entirely for pattern discovery. Across the remaining three blocks, each relationship was tested against a segment of time it had not been measured on — with the training window expanding at each step, so that each successive test used a progressively larger pool of prior history, consistent with how the system accumulates data in production.
A substantial share of the strongest-looking relationships, including several with the largest historical sample sizes, did not survive this process. Relationships that appeared statistically robust across hundreds of historical instances returned results indistinguishable from chance once tested against unseen data — a materially different outcome from a modest reduction in confidence.
A smaller set of relationships retained consistent strength and direction across all three out-of-sample tests. These form the current basis for the system's validated event signals; the remainder are retained for ongoing monitoring, not treated as established findings.
The professional-grade equivalent of this exercise already exists in the form of institutional economic surprise indices — Citigroup's Economic Surprise Index among them — constructed from actual-vs-forecast comparisons weighted by historical market impact. These indices function reliably because their forecast input is a deep, professionally-surveyed consensus across dozens of contributing economists, a data resource not generally available at the retail data-provider level. The concern regarding forecast reliability was directionally correct; the institutional solution addresses it with data most platforms do not have access to.
Separately, most retail "news trading" tools do not attempt to predict release direction at all. The typical implementation places pending orders on both sides of the expected price move and allows realized volatility to determine the outcome, rather than forecasting which direction the market will move.
The resulting approach — measuring the realized price reaction against a genuine no-release baseline, then validating the relationship against data withheld from the discovery process — corresponds closely to "event study" methodology, an established framework in financial economics with a substantial academic literature.
This discipline extends beyond the initial system build. A subsequent refinement — substituting a revised prior-period figure for the original prior-period figure, wherever a revision was available — was proposed on sound reasoning: a revised figure is a more complete piece of information than the figure it replaces. Tested against the same validation framework, the refinement produced weaker results across the system's strongest existing signals and no net improvement across the dataset as a whole. The original, simpler formulation was retained. The result reinforces the operating principle above: a change that is well-reasoned still requires the same empirical validation as the original hypothesis.