catboost versus spectral energy distribution-fitting: estimating galaxy properties under controlled photometric incompleteness

ABSTRACT

Estimating galaxy physical parameters from photometric data is fundamentally challenged by missing measurements that are endemic to astronomical surveys. Using a mock catalogue from the Horizon–AGN hydrodynamical simulation that provides the true physical parameters, we evaluate catboost, a gradient-boosting algorithm that natively handles missing data, under a deliberately adversarial scenario: we train it on 12 photometric bands with increasing levels of injected missingness (10 per cent, 20 per cent, 30 per cent missing per band) and compare its performance against an idealized parametric spectral energy distribution (SED)-fitting reference that uses complete 26-band photometry (no missing data, all bands available). This asymmetric design represents an upper-bound, best-case baseline for traditional methods. Despite this intentional disadvantage, catboost’s performance degradation is limited: moving from complete data to 30 per cent missing, mass RMSE increases from 0.08 to 0.18 dex, SFR RMSE from 0.41 to 0.53 dex, and redshift RMSE from 0.20 to 0.28, while bias remains near zero. Against the ideal SED-fitting reference, catboost trained on only 12 bands with 30 per cent missing values achieves lower errors for mass (0.18 versus 0.28 dex) and SFR (0.53 versus 0.57 dex) and removes the systematic biases present in the SED-fitting results. For redshift, the SED-fitting reference has lower NMAD (0.030 versus 0.093 at extreme missingness) while catboost maintains smaller bias. These results suggest that catboost’s native handling of missing values can offer practical advantages for extracting galaxy properties from imperfect photometric surveys, at least under the conditions explored here.