Bias, Not Skill: Log-Space Air-Quality Forecasts and Cumulative Exposure

Health damage from air pollution comes from accumulated dose, not from any single hour. A week of elevated fine particulate matter matters for an acute stroke trigger. Months to years of exposure matter for the slow narrowing and remodeling of small blood vessels, which the retina lets us photograph directly. Yet air-quality forecasters train and score models on hourly concentration. They usually do so after compressing the heavy right tail of pollution data, either by taking a logarithm of the target or by re-weighting the loss toward rare spikes. This note argues that this mismatch hides a specific, fixable failure. When a model predicts the logarithm of concentration and the prediction is turned back into micrograms with a plain exponential, the result estimates the typical value (the median) rather than the average. It is biased low, and the bias grows with lead time as forecast uncertainty grows. Random errors partly cancel when hours are summed into an exposure window. This bias does not cancel. So for the exposure quantities health actually depends on, the useful horizon of a forecast is set less by how quickly per-hour skill decays than by how quickly this conversion bias builds up.

We propose to measure this effect, to correct it with a back-conversion whose size depends on lead time, and to test whether the correction matters downstream. The downstream test asks whether exposure estimates help explain a person's retinal vascular age, an in vivo marker of systemic and brain blood-vessel health. The deliverable is a benchmark and method for window-integrated, health-relevant pollutant forecasting, on the same reanalysis and station data used by current air-quality models. It comes with a falsifiable biological check, not a claim of causal health effects.

Core Claim

Let $c_{t+h}$ be the concentration $h$ hours ahead, and suppose a model outputs $\hat z_{t+h}$, an estimate of $\log c_{t+h}$, with residual variance $\sigma_h^2$ at lead $h$. If those residuals are roughly Gaussian, the naive conversion $\exp(\hat z_{t+h})$ underestimates the conditional mean by a factor of about $\exp(-\sigma_h^2/2)$. The cumulative exposure over a window of $H$ hours is $E_H=\sum_{h=1}^{H} c_{t+h}$. Its naive estimate therefore carries a relative bias of roughly

$$ b_H \approx \frac{\sum_{h=1}^{H} \mu_h \left(e^{-\sigma_h^2/2}-1\right)}{\sum_{h=1}^{H} \mu_h}, $$

where $\mu_h$ is the true conditional mean at lead $h$. This bias does not shrink as $H$ grows, and it becomes more negative as $\sigma_h$ rises with lead. The random part of the relative error does shrink, roughly like one over the square root of the number of weakly correlated hours in the window.

The claim follows from this. For windowed exposure, log-space models with naive conversion will lose to raw-space or loss-reweighted models beyond some crossover window length, even when they win on hourly error. A lead-dependent smearing correction removes most of that loss. The correction estimates $\sigma_h$ per lead on held-out residuals and rescales each converted prediction. It is the classical smearing estimator, made horizon-aware.

Where This Sits in Prior Work

Two air-quality lines motivate the problem and supply our test beds. AirCast fuses ERA5 weather reanalysis with CAMS EAC4 atmospheric-composition reanalysis in a ClimaX-style vision transformer. It handles skewed particulate data with a frequency-weighted absolute-error loss, reports up to 24-hour leads, and leaves multi-day horizons open. WTMP shows that wavelet decomposition plus a lightweight residual MLP beats Informer-family transformers at 168-hour leads on a single Ruian City station. Neither model asks how its errors combine over an exposure window. Neither compares closed-form target transforms against loss re-weighting. We do both.

Two further papers constrain the design. Predicting Public Health Impacts of Electricity Usage shows that minimizing error on an intermediate physical quantity does not minimize error on the health quantity computed from it. We take that lesson and apply it to the step from concentration to dose. The When and How of Target Variable Transformations finds that quantile-normal and Yeo-Johnson transforms often beat the natural log, and that gradient-boosted trees are largely indifferent to transforms. It studies only static tabular data and scores in a way that does not isolate conversion bias. We therefore include those transforms and trees, and we evaluate on aggregates where conversion bias becomes visible.

What Matters in Deep Learning for Time Series Forecasting? warns that architectural gains often vanish once exogenous covariates are supplied fairly. That warning shapes our baselines: every model receives the same weather inputs.

On the biological side, the retina is well established as a window on systemic and cerebral vessels. RETFound learns transferable retinal representations by masked-image pretraining and predicts 3-year heart failure. DeepRETStroke detects MRI-confirmed silent brain infarction and predicts stroke from fundus photographs. Tran et al. predict incident Parkinson's disease up to 8 years ahead from UK Biobank fundus images. Ludwig et al. and Kim et al. show that retinal "age gaps" track multi-morbidity and survive removal of the vessel tree. All of these treat the eye as a static snapshot with no environmental input. We use them only as the downstream consumer of exposure estimates. We do not build a new retinal model.

Research Program

The first component is a horizon-by-window forecasting benchmark. On the AirCast ERA5 plus CAMS EAC4 MENA setup and on the WTMP Ruian station data, we train each model family under several target treatments:

  • raw-scale squared error;
  • raw-scale frequency-weighted absolute error, as in AirCast;
  • natural log with naive conversion;
  • natural log with lead-dependent smearing;
  • quantile-normal and Yeo-Johnson, each with naive and with simulation-based mean conversion.

We then score every output two ways. The first is hourly error at each lead. The second is error on exposure windows of 24, 72, and 168 hours, built by summing converted hourly predictions. We decompose the window error into a bias term and a variance term, so the mechanism above is observed directly rather than inferred.

The second component is the correction itself. We estimate $\sigma_h$ per lead, and optionally per season and per concentration regime, on a validation year. We then check whether the measured bias curve matches the formula for $b_H$. If residuals are far from Gaussian, we instead use an empirical smearing factor per lead, computed as the mean of exponentiated validation residuals.

The third component is a retinal check. Longer windows, from months up to years, are where vascular remodeling operates, and no forecast has skill there. For those windows, exposure must come from reanalysis history, possibly topped up with a short forecast. We will therefore test the conversion question on reanalysis-derived exposure for UK Biobank participants with fundus imaging. Retinal vascular age gap comes from a RETFound-initialized age regressor. The exposure is multi-year mean $PM_{2.5}$ from CAMS EAC4, computed two ways: by averaging in log space and converting naively, or by averaging in raw space. Either version is assigned at whatever coarse location the approved data access allows. The question is whether the choice of averaging changes the estimated exposure–age-gap association, and whether exposure adds explanatory power beyond age, sex, deprivation, and clinical traits. This is association only. Residual confounding, especially by socioeconomic factors and the cohort-shift problems documented for retinal models, is the main caveat.

Baselines

The naive references are persistence and seasonal climatology. Persistence repeats the last observed value. Climatology predicts the multi-year hour-of-day by day-of-year mean. Both are scored on hourly and window targets using identical splits. Climatology matters most here, because for long windows it may be as good as any learned model.

The numerical reference is the CAMS global forecast. AirCast reports beating it, but the file gives no standalone CAMS number, so no reported result found on Lacuna for CAMS alone.

The strongest learned competitor on the MENA data is AirCast trained with its frequency-weighted loss. The file reports AirCast at $PM_{2.5}$ RMSE 8.82 µg/m³ and $PM_{10}$ RMSE 13.27 µg/m³ for leads up to 24 hours (AirCast). We will first reproduce these figures within a stated tolerance, and only then compare. We will extend its rollout to 168 hours.

On Ruian, the strongest competitor is WTMP, which reports MSE 0.526 at the 720-step look-back and 168-step forecast setting (WTMP). We will reproduce it under the same normalization.

Two further baselines cover the simple-model objections. DLinear and the lobotomized iTransformer from Moretti et al. represent the claim that simple models plus covariates suffice. Gradient-boosted trees on rolling features represent the strong tabular option. A further baseline predicts the window mean directly as its target, in raw space, skipping hourly predictions entirely.

All learned models receive the same ERA5 covariates, the same look-back, and the same hyperparameter search budget, counted in trials and GPU-hours. Target treatment is crossed with architecture, so the effect of the transform is measured within each model family.

The key ablation compares log with naive conversion against log with lead-dependent smearing, using the same trained network. Only the conversion differs, so any gain is attributable to the correction alone. Beating the field means two things. Smeared log-space forecasts must achieve lower window-exposure RMSE than frequency-weighted AirCast and WTMP at 72 and 168 hours, without losing more than a small, pre-registered margin of hourly RMSE at short leads. Their measured window bias must also be statistically indistinguishable from zero where the naive version's is not.

Why not just…?

Why not just keep AirCast's frequency-weighted loss in raw space? It never converts from a transformed scale, so it has no conversion bias, and it already improves $PM_{2.5}$ RMSE by a reported 4.18% (AirCast). This is a real threat. Weighting the loss toward the tail does, however, introduce a different shift in the mean, toward over-prediction in calm periods. That shift would also accumulate over windows. The bias–variance decomposition on window targets settles which method has the smaller net bias at each window length. If frequency weighting wins everywhere, we will report that log transforms should not be used for exposure.

Why not just predict the window-mean exposure directly? Averaging reduces skew, so a model trained on 168-hour means in raw space might avoid the whole issue. It is in the baseline set. It gives up hourly forecasts, which alerting systems need, and needs a separate model per window. If it matches smeared log forecasts on window error, the correct conclusion is that health-oriented systems should forecast windows directly, and we will state that conclusion.

Why not just use gradient-boosted trees, which are said to ignore target transforms? Trees choose the same splits under a monotone transform. Their leaf values, however, are averages in whatever space they are trained in. A tree trained on log targets therefore outputs geometric means, and inherits exactly the conversion bias studied here. Running trees under every target treatment shows whether the "transform-indifference" reported by Nuyts and Davis survives once the scoring is done on summed exposure.

Why not just use climatology for anything beyond a few days? For multi-week and longer windows this may be right. We will report the crossover window at which climatology, or a blend of forecast and climatology, overtakes each learned model. This is also why the retinal analysis relies on reanalysis history rather than forecasts.

Decisive Checks

The mechanism is supported if three things hold together. First, the measured naive-conversion bias on window targets grows with lead in line with the per-lead residual variance, as the formula for $b_H$ predicts. Second, smearing removes most of that bias without hurting hourly error. Third, the ranking of target treatments flips between hourly and window scoring. It is refuted if naive log forecasts show negligible window bias, which would happen if residual variances are small or if the model has implicitly learned to inflate its predictions.

The retinal check is informative in either direction. If the exposure–age-gap association is unchanged by how exposure is averaged, conversion bias is practically irrelevant for chronic endpoints. That would confine the method's value to acute windows of days to a week.