Most of the trading gain from these FX forecasts arrives after Lehman. Filippou, Rapach, Taylor and Zhou report an annual utility gain over optimized carry of 8.46% for their elastic-net model after the break, against 0.89% before it. They forecast monthly dollar moves out of sample using ten country fundamentals whose effects vary with global FX conditions. Both findings appear in the authors' abstract and conclusion. They also say the gains extend throughout the sample. Before September 2008, the advantage is narrow: annual utility gains run from 0.82% to 0.95%, while forecast-driven portfolios have Sharpe ratios of 1.03 to 1.11 against optimized carry's 1.06.

Building the monthly forecast

The authors study ten currencies against the dollar: sterling, the franc, the yen, the Canadian, Australian and New Zealand dollars, the Swedish, Norwegian and Danish currencies, and the euro. Each currency has ten characteristics, measured relative to the US:

Each characteristic is multiplied by six global state variables. Three measure uncertainty and geopolitical risk using newspaper text: US economic policy uncertainty, US monetary policy uncertainty and geopolitical risk. EPU also draws on tax-code and forecaster data. The remaining three measure conditions in FX itself: global volatility (GVOL), illiquidity and correlation. These enter as residuals from AR(1) models fitted using information available at the time. They therefore measure surprises against their own histories, a distinction that matters when interpreting the results.

The pooled regression predicts next month's log spot change from 70 predictors. It has no intercept. A currency's expected drift is therefore zero unless a predictor shifts it.

An elastic net estimates the regression. ERIC, an information criterion, selects the penalty; the authors chose it for its stronger shrinkage than cross-validation. Estimation starts in 1985, with the expanding window refitted each month. The same 70 inputs feed a neural network with four hidden layers containing 16, 8, 4 and 2 neurons. An ensemble takes the average of the two forecasts. Germany, France, Italy and the Netherlands contribute training observations until 1998, then leave the evaluation universe. Out-of-sample evaluation spans 1995:01 to 2020:09, covering 309 months per currency and 248 for the euro, for 3,029 forecasts altogether.

The forecasts then feed a portfolio. A mean-variance dollar investor has risk aversion 5 and uses an EWMA covariance matrix with decay 0.92. Individual currency weights are constrained between -0.5 and 0.5. The bill spread less predicted dollar appreciation gives expected excess return. With the appreciation forecast set to zero, the allocation becomes Bench-Opt, the authors' optimal carry portfolio. Substituting a machine-learning forecast produces ML-Opt.

Which matters more, machine learning or state interactions?

The ensemble's composite out-of-sample R2 is 1.86%, significant at 1% under Clark-West. Every country has a positive result; the UK leads at 3.41%.

The paper's comparisons favor using both shrinkage and interactions. Feeding all 70 predictors into OLS produces a composite R2 of -8.74%; Canada's result alone is -19.75%. With only the ten characteristics and no interactions, the elastic net returns -0.29%. Five individual countries nevertheless have positive, significant results. Adding the interactions to that elastic net lifts R2 to 1.51%. Shrinkage restrains the interaction model's fit to noise, while the interactions supply information worth retaining.

Across the 10 countries, the network achieves 1.74% using those same 70 predictors. Its lead over the elastic net's 1.51% measures the contribution of nonlinearity beyond the linear interactions. We did not find a network run using the ten raw characteristics. Whether the network could discover the state dependence without supplied interaction terms remains open. The authors also caution that composite R2 can include cross-sectional predictability. Individual country results offer a cleaner time-series test: the ensemble's results for Japan (0.19%) and Canada (0.87%) are insignificant, while the other eight countries are significant.

Yield spreads under a volatility surprise

The full-sample elastic net retains 10 of 70 predictors. At 0.17, bill spread times GVOL has the largest coefficient. The standalone bill spread is missing from the retained ten, so in this full-sample fit it influences the exchange-rate forecast entirely through volatility surprises. When FX volatility exceeds its AR(1) prediction, a higher foreign bill yield predicts greater dollar appreciation next month. That appreciation erodes carry. Inflation times GVOL, with a coefficient of 0.09, gives relative PPP the same interpretation. The paper connects the result to Clarida, Davis and Pedersen (2009), whose findings show carry-trade returns falling sharply when FX volatility reaches its highest quartile of observed values.

The authors describe this as a flight to fundamentals. Read the phrase as a description of where the fit is concentrated. These coefficients come from in-sample estimation, though the authors say recursive estimates and the network's variable importance give a broadly similar picture.

Their accuracy-versus-volatility test regresses monthly forecast gains on GVOL in the same month. Pooled slopes are 8.05 for the elastic net, 2.64 for the network and 5.35 for the ensemble; each is significant at 1%. Month-t GVOL is unknown when the month-t forecast is produced. The test thus identifies, after the fact, the months in which forecasts worked. Japan's baseline elastic-net coefficient is negative at -0.88.

The split after Lehman

Over 1995:01 to 2008:08, ML-Opt delivers Sharpe ratios of 1.03 for the elastic net, 1.11 for the network and 1.07 for the ensemble. Bench-Opt records 1.06. Annual utility gains over Bench-Opt are 0.89%, 0.95% and 0.82%, respectively. The paper calls this "significant economic value". Table 3 supplies no test for those utility gains.

During 2008:09 to 2020:09, Bench-Opt's Sharpe drops to an insignificant 0.15. Conventional quintile carry falls to -0.08. The declines in ML-Opt are smaller: the elastic net moves to 0.90 from 1.03, the ensemble to 0.68 from 1.07, and the network to 0.49 from 1.11, significant only at 10%. Utility gains expand to 8.46%, 3.08% and 5.44%. Start with $1 at the end of 1994:12 and the elastic net finishes 2020:09 at $21.76. Bench-Opt finishes at $4.61, plain carry at $2.02.

One regime change accounts for most of the advantage over optimized carry.

The authors maintain that gains occur throughout the sample, citing the late-1990s Asian and LTCM episodes. For allocation, however, the model ranking deserves closer attention. Across the 10 countries, the network's R2 exceeds the elastic net's, 1.74% against 1.51%. Yet its utility gain is less than half as large, 1.98% against 4.50%, and its Sharpe is lower, 0.80 against 0.97.

Mean squared prediction error evaluates each currency against its own realized move. The optimizer instead uses each currency's forecast excess return against the dollar, scales it through the covariance matrix and clips the weight at ±0.5. A plausible interpretation is that better average forecast accuracy can coexist with worse position sizing in the months that determine P&L. For an allocator, the elastic net deserves the study. It is also the model whose workings you can read.

Where the evidence applies

The results hold for a dollar base. The authors report "less predictive power when other currencies serve as the base", specifically the yen and the franc. The universe comprises ten developed currencies. CPI and unemployment use final revised series; the authors disclose this and attribute it to OECD vintages beginning only in 1999. They argue that revisions are small. The three uncertainty indices were constructed post-2016 and applied to archives reaching back to the sample's beginning.

The main portfolio table gives no cost assumption. Supplementary Table IA.19 addresses transaction costs, and Table IA.20 contains the Lustig et al. alphas. Monthly rebalancing and currency weights reaching 0.5 make the cost table consequential.

We did not test this. We have neither FX spot or futures data nor a country-characteristic panel extending back to 1985. Currency ETFs would fail to reproduce spot-plus-carry excess returns for this universe.

The authors report that the advantage survives transaction costs in Table IA.19. The main text gives no cost-adjusted figure, leaving that table as the place to check the post-2008 elastic-net gain.