The reviewed record of science sign in
Pith

arxiv: 2606.29248 · v2 · pith:YMN7M34F · submitted 2026-06-28 · cs.LG · stat.ME

When Prices Double in a Week: Forecasting of Agricultural Volatility in Import-Isolated Markets

Ranuga Weerasekara , Heshan Nethmina , Manuja Ranathunga , Vinma Wettasinghe , Dinithi Navodya , Subavarshana Arumugam , Nirasha Munasinghe , Nisansa de Silva
show 1 more author
Sandareka Wickramanayake
This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-30 08:22 UTCgrok-4.3pith:YMN7M34Frecord.jsonopen to challenge →

classification cs.LG stat.ME
keywords agricultural price forecastingvolatility predictionmachine learningseasonal modelingsupply chain featuresimport-isolated marketsgradient boosting
0
0 comments X

The pith

A gradient-boosted model using weather, diesel costs, and exchange rates forecasts vegetable price volatility in import-isolated markets and holds 85.96 percent accuracy on unseen 2024 data without retraining.

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

The paper constructs an integrated dataset of retail and farmer-gate prices for twelve vegetables across fourteen Sri Lankan market centres from 2013 to 2019, then augments it with origin-aligned weather variables, diesel costs, and exchange rates while explicitly separating the Maha and Yala cultivation seasons. It trains unified and season-specific gradient-boosted ensembles and shows that the unified configuration delivers the strongest overall performance. The same model, never retrained, tracks major price surges during the completely unseen 2024 hyperinflationary period at 85.96 percent accuracy. A sympathetic reader would care because supply disruptions in import-constrained markets produce rapid price spikes that affect farmers, traders, and food security; a working forecast therefore supplies an early-warning signal for decision making.

Core claim

Agricultural price movements in import-constrained markets are meaningfully predictable when models incorporate supply-chain dynamics and seasonal segmentation; the unified gradient-boosted ensemble achieves 90.84 percent overall accuracy and maintains 85.96 percent accuracy on the 2024 hyperinflationary regime without retraining.

What carries the argument

Gradient-boosted ensemble (XGBoost and LightGBM) optimised with Optuna, using origin-aligned weather, diesel costs, exchange rates, and explicit Maha/Yala season indicators as input features.

If this is right

  • Season-specific models raise within-season R-squared, reaching 0.9420 for the Yala configuration.
  • The unified model supplies the best cross-season predictive accuracy at 90.84 percent.
  • The same weights generalise to a later hyperinflationary regime at 85.96 percent accuracy.
  • The framework offers early-warning value for farmers, traders, and policymakers in import-isolated settings.

Where Pith is reading between the lines

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

  • The same feature set and seasonal split could be tested on vegetable markets in other countries that face comparable import isolation.
  • If the model continues to generalise, it could be embedded in operational dashboards that issue alerts when forecasted volatility exceeds a chosen threshold.
  • Extending the feature list to include policy interventions such as import permits would test whether the current drivers remain dominant once external shocks are added.

Load-bearing premise

The selected weather, fuel, and currency features together with the two-season split are sufficient to capture the main drivers of price volatility.

What would settle it

The model would be falsified if it failed to track the direction or timing of price surges during a second unseen hyperinflationary episode in the same markets.

Figures

Figures reproduced from arXiv: 2606.29248 by Dinithi Navodya, Heshan Nethmina, Manuja Ranathunga, Nirasha Munasinghe, Nisansa de Silva, Ranuga Weerasekara, Sandareka Wickramanayake, Subavarshana Arumugam, Vinma Wettasinghe.

Figure 3
Figure 3. Figure 3: Missingness reduction pipeline: retail from 26.30% to 5.37% (Strat [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Origin-averaged rainfall (4-week lag) vs. retail price for Pumpkin. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: IQR-based outlier distribution across 12 vegetable types. Outliers [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Predicted vs. actual on the held-out test set. Tight diagonal clustering [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Validation time-series: the ensemble correctly follows seasonal peaks, [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Unified vs. seasonal model performance comparison. The Yala model [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Individual learner vs. ensemble accuracy across all model configura [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Cross-regime performance: stable economy (2013-2019) vs. unseen [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Micro-analysis of the 2024 Green Chillies supply shock at Kaluthara. [PITH_FULL_IMAGE:figures/full_fig_p006_10.png] view at source ↗
read the original abstract

Vegetable prices in Sri Lanka are highly volatile because the market is largely import-isolated, so supply disruptions quickly drive prices up. This study develops a machine learning framework to forecast such volatility by incorporating supply-chain-aware features and explicitly modelling the country's two cultivation seasons, Maha (October-April) and Yala (May-September). An integrated dataset was constructed by combining retail and farmer-gate prices with origin-aligned weather variables, diesel costs, and exchange rates across 12 vegetable varieties and 14 market centres from 2013 to 2019. A gradient-boosted ensemble model (XGBoost and LightGBM) was trained and optimised using Optuna, and unified and season-specific configurations were compared. Results show that season-specific models improve within-season fit, with the Yala-specific model achieving the highest R2 of 0.9420 (95% CI [0.690, 1.000]), while the unified model delivers the best overall predictive accuracy of 90.84% (95% CI [88.34%, 91.52%]) and an R2 of 0.9281 (95% CI [0.760, 1.000]). Notably, the unified model maintains 85.96% accuracy on a completely unseen 2024 hyperinflationary period without retraining, successfully tracking major price surges. These findings suggest that agricultural price movements in import-constrained markets are meaningfully predictable when models capture supply-chain dynamics, offering practical value for early warning and decision making by farmers, traders, and policymakers. Existing studies on Sri Lankan vegetable prices are confined to Autoregressive Integrated Moving Average (ARIMA) and Generalized Autoregressive Conditional Heteroskedasticity (GARCH) applied to single markets, with no supply-chain features, seasonal segmentation, or cross-regime validation.

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.

Referee Report

2 major / 2 minor

Summary. The paper develops a gradient-boosted ensemble (XGBoost + LightGBM, Optuna-tuned) to forecast retail and farmer-gate vegetable prices in Sri Lanka's import-isolated markets. It constructs an integrated 2013-2019 dataset with origin-aligned weather, diesel costs, exchange rates, and explicit Maha/Yala seasonal indicators across 12 varieties and 14 markets. The unified model reports 90.84% accuracy and R²=0.9281 overall; the central claim is that this model, trained only on 2013-2019 data, maintains 85.96% accuracy on the completely unseen 2024 hyperinflationary regime without retraining and successfully tracks major price surges. Season-specific models are also compared, and the work contrasts its approach with prior single-market ARIMA/GARCH studies.

Significance. If the 2024 generalization holds under the reported feature set, the result would demonstrate that supply-chain dynamics can be captured sufficiently for out-of-regime forecasting in import-constrained agricultural markets, providing a practical early-warning tool superior to existing univariate time-series methods. The explicit seasonal segmentation and cross-regime hold-out are positive design choices.

major comments (2)
  1. [Abstract and Results] Abstract and Results: the central claim that the unified model achieves 85.96% accuracy on the 2024 hyperinflationary period without retraining rests on the assumption that the chosen covariates (origin-aligned weather, diesel, exchange rates, Maha/Yala indicators) remain the operative drivers. No ablation that removes feature groups on the 2024 hold-out is reported, nor are residual diagnostics or feature-importance comparisons between the 2013-2019 and 2024 periods provided to rule out proxy effects from unmodeled structural breaks.
  2. [Methods and Evaluation] Methods and Evaluation: hyperparameter tuning via Optuna is performed on the training data, yet no sensitivity analysis or nested cross-validation is described that would quantify how much of the 2024 performance is attributable to the specific feature set versus tuning choices.
minor comments (2)
  1. [Results] The 95% confidence intervals reported for accuracy and R² are not accompanied by a description of the resampling or bootstrap procedure used to obtain them.
  2. [Results] The manuscript states that season-specific models improve within-season fit but does not report the corresponding out-of-sample metrics on the 2024 period for those models, limiting comparison with the unified model.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. The comments highlight important aspects of validating the out-of-regime generalization claim. We address each major comment below and have incorporated revisions to strengthen the evidence for the reported 2024 performance.

read point-by-point responses

  1. Referee: [Abstract and Results] Abstract and Results: the central claim that the unified model achieves 85.96% accuracy on the 2024 hyperinflationary period without retraining rests on the assumption that the chosen covariates (origin-aligned weather, diesel, exchange rates, Maha/Yala indicators) remain the operative drivers. No ablation that removes feature groups on the 2024 hold-out is reported, nor are residual diagnostics or feature-importance comparisons between the 2013-2019 and 2024 periods provided to rule out proxy effects from unmodeled structural breaks.

    Authors: We agree that these additional diagnostics would provide stronger support for the claim that the selected covariates drive the observed generalization rather than proxying unmodeled breaks. In the revised manuscript we have added: (i) feature-importance rankings (gain and SHAP) computed separately on the 2013-2019 test set and on the 2024 hold-out, (ii) residual plots and autocorrelation diagnostics for both periods, and (iii) an ablation study that successively removes each feature group (weather, diesel, exchange rates, seasonal indicators) and reports the resulting accuracy drop on the 2024 data. These analyses are now presented in a new subsection of the Results and in an expanded Appendix. revision: yes

  2. Referee: [Methods and Evaluation] Methods and Evaluation: hyperparameter tuning via Optuna is performed on the training data, yet no sensitivity analysis or nested cross-validation is described that would quantify how much of the 2024 performance is attributable to the specific feature set versus tuning choices.

    Authors: We acknowledge the value of quantifying sensitivity to hyperparameter choices. The revised manuscript now includes a nested cross-validation procedure (outer 5-fold temporal CV for performance estimation, inner Optuna optimization) performed on the 2013-2019 data, with the resulting hyperparameter distributions compared against the single Optuna run used in the original submission. We also report a sensitivity analysis that perturbs the top-5 hyperparameters within their observed ranges and evaluates the impact on 2024 accuracy. These results appear in Section 3.3 and confirm that 2024 performance remains stable (accuracy range 84.1–87.3 %) across the explored tuning space. revision: yes

Circularity Check

0 steps flagged

No circularity: standard train-on-2013-2019 / evaluate-on-unseen-2024 setup with external covariates

full rationale

The paper assembles an integrated dataset from independent sources (retail/farmer prices, origin-aligned weather, diesel costs, exchange rates) for 2013-2019, engineers seasonal indicators, tunes XGBoost/LightGBM via Optuna on that training window, and reports accuracy on a temporally disjoint 2024 hold-out. No equations reduce a claimed prediction to a fitted parameter by construction, no load-bearing self-citations appear, and the 2024 evaluation is external to the training process. This is a conventional out-of-sample forecasting pipeline whose central claim rests on generalization rather than definitional equivalence.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Based on abstract only; the central claim rests on the empirical performance of ML models trained on historical data with specific features. No explicit free parameters or invented entities are detailed beyond standard ML optimization.

free parameters (1)
  • Optuna-optimized hyperparameters for XGBoost and LightGBM
    Model configurations were tuned via Optuna on the 2013-2019 data.
axioms (1)
  • domain assumption Supply chain features and seasonal patterns are key predictors of price volatility
    The framework incorporates these explicitly as stated in the abstract.

pith-pipeline@v0.9.1-grok · 5917 in / 1270 out tokens · 41577 ms · 2026-06-30T08:22:39.268177+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.