Laguerre's market-timing strategy earns its headline certainty equivalent against a long-only benchmark that holds Treasury bills. The ridge penalty is the paper's more useful result.
Building the forecast
Random features ridge regression (RFRR) feeds predictors through hundreds of random cosine transformations, then fits ridge regression to the resulting features. Kelly and coauthors (2024) argued that adding features and taking the penalty toward zero improves market timing. Laguerre asks when RFRR can beat a forecast of zero excess return.
He calls the answer the SNR frontier: for each ridge penalty λ, it gives the minimum signal-to-noise ratio needed to beat zero. A trivially biased estimator cannot clear that frontier at any SNR. The paper then introduces posterior drift, meaning out-of-sample coefficients equal w times their in-sample values. Signal at risk (SaR) measures the share of signal an investor expects to lose. As drift raises the frontier, investors expecting greater losses need more shrinkage. Under isotropic true coefficients, with signal evenly distributed across directions, Proposition 3 proves that RFRR is always trivially biased against the zero forecast and can never beat it.
The empirical target is the monthly CRSP value-weighted return from 1930 to 2020, using the Goyal, Welch and Zafirov (2024) data. The model has 15 predictors, comprising 14 macro indicators and the lagged return. A generalized method of moments (GMM) fit estimates signal at τ = 2×10^-5 and the alignment coefficient at r = 0.35. Its Sargan-Hansen J is 0.65, below the 5.99 critical value. With monthly return variance of 2.94×10^-3, the in-sample SNR comes to 0.68%.
Trading runs monthly from 1990 to 2020. Each fit uses 600 random features, a 10-year window and a penalty of 1.2×λ_min. The strategy averages 500 forecasts before choosing a Markowitz market/T-bill weight. With risk aversion 2 and long-only limits, it earns an annualized certainty equivalent of 7.95%; the zero-forecast rule earns 2.63% (p = 0.03%).
What does a fading signal cost?
Without drift, λ_min is about 1.3×10^-3. Protection against a 75% signal loss raises it to 1.9×10^-3. At a 99.99% expected loss, it reaches roughly 0.14, about a hundred times its no-drift level. Simulations follow the same pattern. For SNR of 10% and w = 0.7, meaning about half the signal is lost, the required penalty is at least 4. Without drift it is 1.2.
These frontiers describe average excess error under i.i.d. draws and a power-law alignment assumption. They carry no guarantee for trading. The author describes his SNR estimation as "basic" and uses unconditional return variance as a proxy for noise. Section 7.1 also draws its figures from the full 1930 to 2020 sample. Trading recalibrates on an expanding window that starts with 1930 to 1989. The headline penalties therefore describe the full period; they do not tell us what a trader in 1990 would have selected.
The penalty floor is fairly stable. Across nine choices of window and bandwidth, λ_min at w = 0.01 ranges from 0.127 to 0.216. Estimated SNR varies from 0.21% to 1.5%, and r̂ becomes negative (−0.24) at T = 60, γ = 2.5. J statistics reject in three specifications (72.11, 84.14, 124.24). These are shaky calibrations, yet the floor changes by at most a factor of 1.7 (0.127 to 0.216). There is also a discrepancy within the paper: the main text reports r̂ of 0.35, whereas the sensitivity table gives 0.27 for the same T = 120, γ = 2 setting.
More features increase the drift penalty
The drift model rescales every coefficient by the same factor: β_oos = wβ_is. The paper says an independent random shock to out-of-sample coefficients leaves the deterministic equivalent unchanged, with the error bound absorbing the shock. A systematic move of signal into different eigendirections remains outside the model.
Within that rescaling model, effective regularization decreases as the feature count p increases. The drift term is largest at low effective regularization. Consequently, the gap between drifted and in-distribution error grows with p. The result establishes that widening gap; it makes no claim that total error always rises with p. Adding random features enlarges the drift penalty and offers no insurance against a fading signal.
7.95% certainty equivalent against 2.63% in T-bills
The zero-return forecast produces a Markowitz weight of −r_f/(aσ²). A long-only [0,1] constraint clips that weight to zero, leaving the benchmark entirely in T-bills. Its returns are therefore 2.63%, 2.62% and 2.62% at the three risk aversions. With the [−1,2] constraint, the rule shorts the market and earns −1.58% at a = 2. The paper calls the zero forecast "a short-term reversal strategy". Yet the rule uses no past returns. The setup's only reversal ingredient is the lagged return among RFRR's predictors.
The benchmark choice is explicit. The author cites Gu et al. (2020), who find that historical means fail to beat zero for individual stocks. He concludes that zero makes the comparison "more demanding and meaningful". That defence addresses forecast error. In the Markowitz allocation with a [0,1] constraint, zero instead produces 100% T-bills (2.63%). The author acknowledges that the historical-mean comparison is "more nuanced" for aggregate indices and retains zero.
We did not find comparisons against buy-and-hold, a historical mean or a ridgeless model. Nor did we find a conventionally tuned ridge model, the comparison needed to establish whether 1.2×λ_min contributes anything. We found no transaction costs, Sharpe ratio or turnover. Certainty equivalent is the only reported return measure. We also did not find which SaR Table 1 uses.
The highest figure is 11.75%, under the [0,2] constraint, which permits market exposure up to twice the endowment. The text describes that case as limited to 100% leverage, while Table 1 lists it under [0,2]. For single stocks, the reported mean is 5.11% against 0.55%, and 89% of portfolios finish ahead at a = 2. The 355 names were chosen for having return data throughout January 2000 through December 2019. This is a survivor universe. The test itself covers 2011 to 2019.
Why we did not trade the strategy on SPY
We could not run it.
Our data lack the Goyal, Welch and Zafirov predictor panel. Replacing it with a handful of FRED series would amount to testing a different forecasting model. SPY could supply aggregate US equity exposure in place of the CRSP index. That would implement the equity leg without reproducing the paper's portfolio returns.
The discipline transfers: choose the ridge penalty using an estimated 0.68% SNR and an explicit signal-loss assumption, then stop interpolating. This agrees with Shen and Xiu's result that ridge needs non-vanishing shrinkage to beat zero. We would change our view of the trading claim if the 1.2×λ_min rule beat cross-validated ridge over the same 1990 to 2020 window.