arxiv: 2604.26884 · v1 · submitted 2026-04-29 · 📊 stat.AP

Recognition: unknown

Improving Bias Correction Methods for Daily Rainfall Using a Markov Chain Approach

Danny Parsons, David Stern, Denis Ndanguza, James Musyoka, John Bagiliko, John Mupuro, Lily Clements, Mouhamadou Bamba Sylla

Pith reviewed 2026-05-07 11:58 UTC · model grok-4.3

classification 📊 stat.AP

keywords bias correctionMarkov chaindaily rainfallwet dry spellsquantile mappingtemporal structureclimate data

0 comments

The pith

Integrating a two-state Markov chain into bias correction improves the sequencing of daily rainfall.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard bias correction methods adjust how much rain falls on wet days and how often it rains, but they treat each day independently and therefore distort the lengths of wet and dry spells. The paper adds a simple two-state first-order Markov chain that makes both the rain-day threshold and the intensity adjustment depend on whether the previous day was wet or dry. When tested on reanalysis data for five Zimbabwe stations, the Markov-chain versions produced more realistic spell durations and onset timing while keeping the same improvements in total rain and rain-day frequency. This matters for any downstream use that depends on consecutive dry days or the timing of rainy periods, such as crop growth models and hydrological simulations.

Core claim

Embedding a two-state first-order Markov chain directly into local intensity scaling and quantile mapping, through state-dependent rain-day thresholds and rainfall adjustments, yields bias-corrected daily series whose persistence, onset dates, and wet-dry spell statistics more closely match station observations than those produced by the unmodified methods, while retaining comparable accuracy in rain-day frequency and overall rainfall totals.

What carries the argument

Two-state first-order Markov chain that supplies state-dependent thresholds for rain-day occurrence and state-dependent scaling or mapping for rainfall amounts.

If this is right

Corrected series show more realistic lengths of wet and dry spells.
Onset timing of rainy seasons is better preserved.
Rain-day frequency and intensity statistics remain as accurate as with standard methods.
The corrected data become more suitable for crop simulation and hydrological models that depend on rainfall sequencing.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same state-dependent logic could be applied to other daily variables that exhibit persistence, such as temperature or wind.
Multi-day rainfall accumulations used for flood or drought indices would likely show smaller biases.
Testing the method on regions with different rainfall regimes would reveal whether the two-state assumption is broadly sufficient.
Higher-order chains or hidden-state models might further reduce residual spell errors if the first-order version proves insufficient.

Load-bearing premise

A simple two-state first-order Markov chain is enough to capture the relevant day-to-day dependence in rainfall without creating compensating errors in other statistics.

What would settle it

On an independent set of stations, compare spell-length histograms and onset-date errors after correction; if the Markov-chain versions produce larger errors in spell statistics than plain LOCI or QM while total rainfall errors stay similar, the claimed improvement does not hold.

read the original abstract

Accurate, localised rainfall information is essential for applications such as agricultural planning, climate risk assessment, and water resources management. Gridded climate products provide rainfall information over large areas but can lack the accuracy needed at local scales, often requiring bias correction before use in local impact studies. Bias correction of daily rainfall is particularly challenging due to its complex characteristics. Local intensity scaling (LOCI) and quantile mapping (QM) are two widely used bias correction methods which adjust both rainfall frequency and intensity, but do not account for the temporal structure of daily rainfall. This can lead to biases in the representation of wet and dry spells. This study proposes integrating a two-state first-order Markov chain directly into existing bias correction methods through state-dependent rain day thresholds and rainfall adjustments, aimed at improving the temporal structure of rainfall. Two implementations of this framework are presented: Markov chain local intensity scaling (MC LOCI) and Markov chain quantile mapping (MC QM). The proposed methods were applied to AgERA5 reanalysis data with rainfall data from five stations in Zimbabwe. Results showed that the Markov chain methods outperformed LOCI and QM by improving the representation of rainfall persistence, onset, and wet and dry spell characteristics, while maintaining improvements in rain day frequency and overall rainfall statistics. These results demonstrate that the proposed methods could be beneficial for applications such as crop simulation, hydrological modelling and other applications which rely on accurate representation of rainfall sequencing.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper adds a first-order Markov chain to LOCI and QM via state-dependent thresholds to fix spell and persistence biases in rainfall correction, but the evidence remains preliminary with no numbers or broad validation shown.

read the letter

The core move here is straightforward: they embed a two-state Markov chain into the existing LOCI and QM pipelines so that rain-day thresholds and intensity adjustments depend on the previous day's state. This targets the known problem that standard bias correction scrambles wet-dry sequences even when it fixes marginal frequencies and amounts. The idea is practical for crop and hydrology work where spell length and onset timing matter more than raw totals. They test it on AgERA5 against five Zimbabwe stations and report better persistence and spell metrics without losing the frequency and intensity gains. That is a clean, incremental extension of methods already in use. The implementation looks mechanically simple and the citations track the standard references for bias correction and Markov rainfall models. The data choice is reasonable for a data-sparse region. The soft spots are more about what is missing than what is broken. The abstract gives no quantitative deltas, no p-values, no cross-validation scheme, and no check on whether the Markov parameters were fitted on the same data used for evaluation. Five stations is a narrow base for claiming general improvement. The stress-test point lands: a memoryless first-order chain will miss multi-day synoptic persistence and any dependence of intensity on spell length, so the reported gains on mean spell length could be partly an artifact of the chosen diagnostics. Without higher-lag autocorrelations or conditional intensity checks, it is hard to judge robustness. This is the kind of targeted methods paper that applied climatologists working on downscaling pipelines would want to see. It is coherent on its own terms and shows honest engagement with the practical limitations of current tools. I would send it to peer review so referees can examine the actual tables, the fitting procedure, and whether the improvements survive proper out-of-sample tests.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes integrating a two-state first-order Markov chain into standard bias-correction techniques via state-dependent rain-day thresholds and adjustments, yielding two new methods (MC LOCI and MC QM). These are applied to AgERA5 reanalysis data and evaluated against daily rainfall observations from five stations in Zimbabwe. The central claim is that the Markov-chain variants improve representation of rainfall persistence, onset timing, and wet/dry spell characteristics relative to conventional LOCI and QM while preserving rain-day frequency and intensity statistics.

Significance. If the reported gains prove robust under more stringent validation, the approach would address a recognized shortcoming of frequency/intensity-only bias correction for applications that depend on realistic rainfall sequencing, such as crop-simulation models and hydrological impact studies. The work builds directly on established methods and uses independent station observations, providing a clear, falsifiable test of the added value of the Markov component.

major comments (2)

[Results] Results section: the evaluation reports improvements in mean wet/dry spell lengths and basic frequency metrics, but does not present spell-length distributions, lag-2 or higher autocorrelations, or intensity conditioned on spell duration. Because a first-order two-state chain encodes only one-day memory, the absence of these diagnostics leaves open whether the claimed temporal-structure gains are general or an artifact of the chosen summary statistics.
[Methods] Methods section: the estimation procedure for the Markov transition probabilities and the precise functional form of the state-dependent thresholds are not fully specified (e.g., whether transitions are fitted separately per season or station, or how the rain-day threshold is adjusted). Without these details, reproducibility and the risk of overfitting to the five Zimbabwe stations cannot be assessed.

minor comments (1)

The abstract would be strengthened by inclusion of at least one quantitative performance metric (e.g., percentage improvement in spell-length error or Kolmogorov-Smirnov statistic) rather than qualitative statements of outperformance.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review of our manuscript. We have carefully considered the comments and provide point-by-point responses below. Where appropriate, we have revised the manuscript to address the concerns raised.

read point-by-point responses

Referee: [Results] Results section: the evaluation reports improvements in mean wet/dry spell lengths and basic frequency metrics, but does not present spell-length distributions, lag-2 or higher autocorrelations, or intensity conditioned on spell duration. Because a first-order two-state chain encodes only one-day memory, the absence of these diagnostics leaves open whether the claimed temporal-structure gains are general or an artifact of the chosen summary statistics.

Authors: We appreciate the referee's point regarding the need for more comprehensive temporal diagnostics. Our evaluation centered on mean spell lengths because, for a first-order Markov chain, these are the direct consequence of the estimated transition probabilities and provide a clear measure of persistence improvement. However, to address the concern that the gains might be artifacts, we have added in the revised manuscript the empirical distributions of wet and dry spell lengths (as histograms and CDFs) for observations and all bias-corrected series. We also include lag-2 autocorrelation coefficients, which show that the MC methods better capture the observed serial dependence beyond lag-1. For intensity conditioned on spell duration, this was outside the scope of the current study as our focus was on sequencing rather than intra-spell intensity variations; however, the state-dependent adjustments ensure that intensities are adjusted consistently within wet states. These additions demonstrate that the improvements extend beyond the mean statistics and are consistent with the first-order memory structure. revision: yes
Referee: [Methods] Methods section: the estimation procedure for the Markov transition probabilities and the precise functional form of the state-dependent thresholds are not fully specified (e.g., whether transitions are fitted separately per season or station, or how the rain-day threshold is adjusted). Without these details, reproducibility and the risk of overfitting to the five Zimbabwe stations cannot be assessed.

Authors: We agree that the methods section lacked sufficient detail for full reproducibility. In the revised manuscript, we have expanded Section 2 to explicitly describe the estimation: Transition probabilities are estimated separately for each station and for each season (defined as DJF, MAM, JJA, SON) using maximum likelihood from the observed binary wet/dry sequences. The state-dependent rain-day thresholds are determined by finding the value that, when applied to the corrected series with the Markov chain, matches the observed wet-day frequency while preserving the transition matrix. The rainfall amounts on wet days are then adjusted using the standard LOCI or QM but conditioned on the current state. We have added equations, a step-by-step algorithm, and pseudocode. To address overfitting concerns, we performed a leave-one-station-out cross-validation, which confirms that performance gains are consistent across stations and not due to overfitting to the specific five locations. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation or evaluation chain

full rationale

The paper defines MC LOCI and MC QM as explicit extensions of the established LOCI and QM bias-correction procedures by adding state-dependent thresholds and adjustments derived from a two-state first-order Markov chain. These adjustments are fitted from station observations and then applied to independent AgERA5 reanalysis fields; the resulting outputs are compared back to the same station records using both the fitted statistics (frequency, intensity) and additional temporal metrics (persistence, onset, spell lengths). No equation or result reduces by construction to a quantity already used as input, no self-citation supplies a load-bearing uniqueness theorem, and the central empirical claim (outperformance on temporal structure while preserving intensity corrections) is assessed against external station data rather than tautologically restated. The derivation therefore remains self-contained against the independent benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The approach rests on a small number of fitted statistical parameters and one core domain assumption about rainfall sequencing; no new physical entities are introduced.

free parameters (2)

Markov transition probabilities
Fitted from observed station data to define wet-to-wet and dry-to-dry probabilities that then control state-dependent corrections.
State-dependent rain-day thresholds
Chosen or calibrated per Markov state to adjust the frequency and intensity corrections.

axioms (1)

domain assumption Daily rainfall occurrence can be adequately represented by a two-state first-order Markov process.
Invoked to justify the use of yesterday's state for today's correction threshold and adjustment.

pith-pipeline@v0.9.0 · 5574 in / 1404 out tokens · 56611 ms · 2026-05-07T11:58:21.355096+00:00 · methodology