Building a Macro-to-FX Framework: How Institutions Turn Data Into Currency Bets

For decades, the consensus in academic economics was bleak: no model of exchange rates could reliably beat a random walk in out-of-sample forecasting. The Meese-Rogoff (1983) finding was treated as definitive, and for a generation of economists it was. Fundamental models were dismissed. The exchange rate was thought to be essentially unforecastable from observable macro variables.

This consensus has been overturned. A combination of better models, better data, and a crucial theoretical insight from Engel and Wu (2023) has demonstrated that macro fundamentals do explain exchange rates — systematically, out of sample, and in real time. The implication for practitioners: a macro-to-FX framework built on the right transmission channels is not just theoretically justified, it is empirically validated.

Why Exchange Rate Models Failed for 40 Years

The Meese-Rogoff finding had a specific and often misunderstood character. It showed that models based on current macro fundamentals could not beat the random walk. But exchange rates are forward-looking — they respond to expected future fundamentals, not current ones. Models that used current data to predict future exchange rates were therefore systematically misspecified.

The second problem was data quality. The macro variables that theory says should drive exchange rates — relative money supplies, output gaps, inflation differentials — were measured with substantial noise and revised significantly after initial release. Models estimated on revised data and tested on real-time data performed worse than the random walk, but the culprit was data revision, not theory failure.

A third issue was linearity. Most early models assumed linear relationships between macro variables and exchange rates. The actual relationships, as the volatility regime research demonstrates, are highly nonlinear — the same macro signal has different effects depending on the risk environment, the position of the carry trade cycle, and the state of global liquidity.

The Engel-Wu Breakthrough

Charles Engel and Steve Wu's 2023 paper, "Forecasting the US Dollar in the 21st Century," provided the most compelling recent evidence that exchange rate models work. Their key insight: exchange rates embed a risk premium that is large, time-varying, and predictable from observable macro variables.

When this risk premium is explicitly modeled — rather than treated as noise — macro-fundamental models substantially outperform the random walk at all horizons from one month to several years. The variables that best predict the risk premium are exactly those that the transmission chain framework highlights: rate differentials, volatility regime indicators, and global risk appetite measures.

Four Transmission Channels

A complete macro-to-FX framework integrates four distinct transmission channels, each operating on a different horizon and through a different mechanism:

Channel 1 — Volatility Regime (Gating Layer): Bond volatility and equity volatility together define the macro risk environment. This layer does not directly produce a directional FX bias; instead, it gates the output of all subsequent layers. High volatility suppresses the reliability of fundamental signals and caps the allowable conviction level.

Channel 2 — Rate Structure (Directional Layer): The 2-year rate differential is the primary directional input for the dollar against major pairs. Yield curve spread configurations (3m2s and 2s10s) provide near-term and structural horizon signals respectively. This is the layer that most directly connects central bank policy to exchange rate direction.

Channel 3 — Dollar Directional Bias (Aggregation Layer): The dollar's overall directional bias — incorporating the Dollar Smile framework, DXY-equivalent signals, and bilateral rate differentials — is aggregated from the rate structure layer. This layer identifies whether the dollar is in a structurally bullish, bearish, or sideways configuration.

Channel 4 — Pair Bias (Application Layer): Individual pair biases are derived by applying the dollar directional bias to each pair's specific characteristics: its safe-haven properties (JPY, CHF), its energy sensitivity (EUR), its commodity exposure (CAD, AUD, NZD), or its carry position (EM pairs). This layer produces the actionable output: which pairs have the most reliable and highest-conviction directional bias.

Conviction Scoring and Regime Gating

The most important structural innovation in modern institutional macro-FX frameworks is conviction scoring that is explicitly conditioned on the volatility regime. This is not just about position sizing — it is about recognizing that the signal-to-noise ratio of fundamental inputs varies systematically with the macro environment.

In a calm volatility regime, rate differentials are highly predictive, carry trades function normally, and fundamental signals can be acted on at full conviction. In a stressed regime, fundamental signals degrade rapidly — safe-haven flows, forced deleveraging, and liquidity spirals overwhelm fundamentals. Full conviction in stressed conditions is not confidence; it is recklessness.

The 4xForecaster conviction scale (1–4 dots) implements this regime conditioning explicitly. A maximum 4-dot conviction is only possible in a Calm regime. In Caution, conviction is capped at 3. In Stress, no directional bias is presented at maximum conviction — the framework acknowledges the reduced reliability of all fundamental inputs.

Cross-Asset Confirmation as the Final Gate

Before any directional bias is presented at full conviction, cross-asset confirmation is checked. Sector rotation (XLU/SPY, XLE, XLF relative performance) must be consistent with the macro FX thesis. When cross-asset signals align with the fundamental chain, conviction is maintained. When they contradict — defensive rotation while rate differentials favor the dollar, for example — conviction is reduced until the discrepancy resolves.

This cross-asset gate is what distinguishes a systematic macro framework from a simple rate-differential model. The framework is never fully committed to any single signal; it continuously integrates multiple independent information sources and adjusts conviction accordingly.