The methodological advance over recurrent MOCA is insufficiently established, and the theoretical results add limited technical novelty. MOCA already maintains predictive states for different run lengths, updates their probabilities using predictive likelihoods and a hazard transition, and marginalizes their predictions. Crucially, it also includes an explicit LSTM-based implementation; recurrent states under shared learned parameters are therefore already part of that framework ([1], sections 4.1-4.3). Its training procedure differentiates through the filtering computation.
CHARM’s apparent differences are its particular autoregressive GRU and mixture-density implementation, separate crop-based predictor training, and forecasting evaluation. These may be useful choices, but the paper does not sufficiently establish their significance relative to the existing framework. The SH-MOCA baseline does not resolve this concern: it uses a different shared-feature Bayesian predictor, so its comparison with CHARM changes the predictive model as well as the method under evaluation. A comparison with a matched recurrent implementation is especially important here.
The theory provides interpretation but does little to establish an additional contribution. Lemma 1 is the standard population log-loss optimality argument under realizability. Proposition 1 assumes exact candidate conditionals and a matching changepoint model. Proposition 2 follows from the KL chain rule and marginalization. These statements appear reasonable, but they neither control the errors of the learned, capped implementation nor establish when CHARM improves over simpler context policies. Moreover, section 3.2 explicitly acknowledges that the conditionals learned from training crops need not match those assumed in Proposition 1.
The presentation makes this contribution boundary harder to assess. Although the combination is specified mathematically in section 3, the paper would benefit from a direct explanation of what is inherited from MOCA, what changes, and why the changed training procedure should produce useful likelihoods for comparing candidate histories. A self-contained algorithm description or a data flowing image would help in the readability and overall understanding of the method.
The overall ranking in Table 1 is more favorable than the evidence attributable specifically to CHARM’s context mechanism. In Table 2, on Binance, CHARM and GRU-global achieve essentially the same NLL. Thus, the Binance results do not establish an improvement from adding context inference to the recurrent predictor, leaving the clearest evidence for the proposed mechanism concentrated on Oxford-Man.
The limited number of training seeds is a substantial concern: the real-data experiments use only two neural training seeds, and the synthetic experiments use one. The reported bootstrap intervals are conditional on these fitted models and therefore do not establish robustness to initialization and training randomness. Evaluation with more independent training seeds is necessary, together with seed-level paired comparisons and measures of variability, to determine whether the observed advantages are reproducible across trained predictors.
The evaluation also covers a single forecasting task (daily log realized volatility) across four equity indices and two cryptocurrency pairs. Additional datasets, preferably covering other nonstationary forecasting settings, are necessary to establish whether the method’s incremental benefit extends beyond Oxford-Man. Given the close relationship to MOCA, the paper needs stronger empirical evidence of a consequential and reproducible advantage. The current combination of limited training replication and narrow dataset coverage does not sufficiently establish that advantage.
As a minor presentation comment (which does not affect my recommendation), I find the opening quotation unnecessary: it adds little to the scientific motivation and makes the introduction less direct. I would prefer the paper to begin with a precise statement of the forecasting problem, the limitations of existing approaches, and the specific gap that CHARM addresses.
Forecasting Beyond Finance and the Cost of Context Marginalization
Thank you for asking when context marginalization justifies its cost. We added two hourly forecasting tasks outside finance, a held-out validation diagnostic, and benchmarks across four run-length caps (§5.6 and the appendix).
Q1. Do the gains extend beyond financial volatility?
For next-hour Australian electricity demand, the equal-state average favored CHARM over GRU-global in all five seeds (mean NLL difference −0.03002, adjusted interval [−0.03951, −0.02046]). South Australia accounted for 91% of that net gain and also favored CHARM over tuned static (−0.02856, adjusted interval [−0.04495, −0.01053]). Across all five states, the gain over tuned static was uncertain (−0.00327, adjusted interval [−0.00746, +0.00166]). On both hourly transformer oil-temperature series, ETTh1 and ETTh2, both controls had lower NLL than CHARM in every seed.
Q2. When is the additional inference cost worthwhile?
We first checked whether different histories produced different forecasts. On validation targets supporting the long-history comparison, the mean CDF discrepancy between the 128-step candidate and GRU-global was 0.0221 on Oxford-Man and 0.000075 on Binance. The latter comparison covered 648 of 1,880 targets. A disagreement-based switch selected on early validation still required the candidate forecasts and did not improve on always using CHARM on later validation. On that later split, CHARM minus tuned static NLL was −0.01484 on Oxford-Man and −0.00045 on Binance.
For deployment, we compare CHARM, global, and tuned static by paired NLL on held-out validation dates and choose the best-scoring policy that meets the latency budget. The table gives median milliseconds per forecast on the same CPU, using two threads, batch size one, and five repetitions on one representative checkpoint per dataset.
Increasing the cap from 128 to 256 raised latency on both datasets. In the original two-seed test evaluation, it changed mean NLL by less than 0.0001 on either dataset.