PaperlayerView exact revision on arXiv

Source-linked AI summary

Forecasting Weather-Driven Price Dynamics Across Sri Lankan Tea Market Catalogues

Hesandi Mallawarachchi, Senilka Madurapperumage, Nadil Kulathunge, Thilokya Angeesa, Nethsith Gunaweera, Sandeepa Weerasekara, Patalee Narasinghe, Nisansa de Silva, Sandareka Wickramanayake

arXiv:2608.24894v1econ.GNcs.LG

TL;DR

The study addresses limited evidence on how localized weather affects weekly prices across Sri Lankan tea catalogues. It constructs a weather-enriched auction dataset and applies Granger causality with tree-based models, finding catalogue-specific weather sensitivities and tailored model performance.

  • Problem

    Prior work has generally pooled Sri Lankan tea grades or used annual aggregates, leaving catalogue-level weather effects on weekly Colombo Tea Auction prices insufficiently tested.

  • Method

    The study combines 105 weekly broker reports with region-specific weather data, lagged features, Granger causality analysis, and four tree-based machine-learning models.

  • Results

    Catalogue-specific modeling is heterogeneous: LightGBM performs best for High Grown, Low Grown, and Off-Grade, while Random Forest performs best for Dust; unified LightGBM reaches RMSE = 137.95 and R² = 0.9515.

  • Takeaways & Limitations

    Forecasting should account for localized weather patterns and catalogue-level differences, with real-time auction-day signals potentially improving forecasts for High Grown and Dust.

Abstract

from arXiv · show

The Colombo Tea Auction (CTA) plays a vital role in determining global tea prices, yet the relationship between local weather conditions and price behavior across different tea catalogues has not been thoroughly explored. In this study, we develop a novel, structured dataset by extracting information from 105 weekly broker reports spanning late 2023 to 2026, and combined with region-specific weather data. Our analysis focuses on four main tea catalogues of Sri Lankan tea: High Grown, Low Grown, Off-Grade, and Dust. To better understand the factors influencing tea prices, we apply Granger causality analysis alongside tree-based machine learning models: Random Forest, XGBoost, LightGBM, and Gradient Boosting. Our results show that while market dynamics are primary drivers, weather conditions also have significant effects. Notably, Low Grown tea shows strong sensitivity to precipitation and sunshine duration (p<0.05) across 1-3-week lags. Off-Grade and Dust catalogues also exhibit significant responses to temperature variations. Catalogue-specific modelling outperformed unified approaches, with LightGBM emerging as the superior model for three out of four catalogues. Overall, this study highlights the importance of considering both localized weather patterns and catalogue-level differences when forecasting tea prices, offering a more precise and practical framework for the tea industry.

I. INTRODUCTION

Sri Lankan tea is economically important, and the study frames its prices through geographical differences in elevation and growing conditions. The Colombo Tea Auction provides the weekly market context for analyzing these distinctions.

  • Tea contributes roughly 10% of Sri Lanka’s agricultural export earnings and supports more than two million livelihoods.
  • The Colombo Tea Auction is Sri Lanka’s primary weekly price-discovery mechanism for Ceylon tea.Its price benchmarks flow into forward contracts and retail shelves in more than 90 countries.
  • Elevation strongly influences tea quality, flavor profile, and growing conditions.
  • High Grown teas are cultivated above approximately 1,200 m and include regions such as Nuwara Eliya, Western High, Uda Pussellawa, and Uva.
  • Medium Grown teas come from roughly 600–1,200 m and are often grouped with High Grown teas because of similar characteristics and market positioning.
  • Low Grown teas are cultivated below approximately 600 m, mainly in southern and southwestern regions including Galle, Matara, and Ratnapura.They are known for stronger liquors with good color and body.

2) Categorization by Catalogue Type:

Sri Lankan auction tea is divided into four distinct catalogues with different physical characteristics and price levels. Existing data and methods have often pooled these categories, motivating catalogue-specific analysis and a structured weather-enriched dataset.

  • The four catalogues are High Grown, Low Grown, Off-Grade, and Dust.They differ in how tea is sorted, processed, presented, and priced at the Colombo Tea Auction.
  • Off-Grade and Dust include teas from High, Medium, and Low Grown regions, unlike geographically defined elevation groupings.
  • The catalogues show substantial differences in average auction prices, but prior computational work mostly pooled grades or used annual aggregates.
  • CTA price information is distributed across weekly broker PDFs with semi-structured tables and no public machine-readable archive.
  • The study addresses this gap with a machine-parsed CTA corpus, region-specific weather alignment, and formal tests of weather effects on catalogue-level prices.
  • Catalogue-specific price drivers can support more accurate broker advice and better sourcing and bidding decisions.

II. RELATED WORK

Prior Sri Lankan tea-price studies emphasize macroeconomic or aggregated outcomes, while limited work addresses weekly auction dynamics and lagged weather effects across catalogues. This study combines broker-report data with region-specific meteorological observations to examine that gap.

  • Earlier Sri Lankan tea-price studies mainly examine exchange rates, oil prices, foreign supply, and aggregate export prices at monthly or annual frequency.
  • Those macroeconomic analyses pool grades and therefore cannot resolve the weekly catalogue-level dynamics used by brokers and buyers.
  • Limited research has examined weekly Sri Lankan tea auction prices using time-series techniques for lagged weather effects across market catalogues.
  • Agronomic studies link rainfall, temperature, and sunshine to tea yield and leaf quality, but these relationships have not been directly tested against weekly catalogue-level auction price formation.
  • The study uses 105 weekly Forbes & Walker broker reports collected from November 2023 to March 2026.
  • Weather variables from Open-Meteo include precipitation, mean temperature, and sunshine duration for High, Medium, and Low Grown regions.

IV. METHODOLOGY

The methodology builds a modeling-ready dataset by extracting weekly auction information, aligning report-based and regional weather features, and constructing lagged predictors. It then organizes the data for catalogue-level price analysis.

  • The methodology follows a five-stage end-to-end pipeline for Colombo Tea Auction price modeling.
  • Structured data is extracted from weekly Forbes & Walker PDF reports using pdfplumber and a custom multitable pipeline.
  • The dataset overview is presented in Table I.
  • The extraction process generates nine standardized CSV files per run.
  • Weather data combines report-text conditions and crop descriptions with daily Open-Meteo observations aggregated into regional 7-day windows.
  • Lagged weather features at 1, 2, and 3 weeks before each auction date capture delayed supply effects.The reference date is auction_date −7×lag.
  • The master dataset joins price, sale-level context, and region-aligned weather using sale_id, producing 12,233 modeling-ready rows.The target price_mid_lkr is calculated from reported lower and upper price bounds, with a lower-bound fallback.

1) Temporal lags:

The study constructs lagged weather and structural price features to model weekly tea-auction dynamics, while exploratory results show clear catalogue segmentation and uneven data quality.

  • Temporal lags:: Weather lags cover precipitation, temperature, and sunshine duration measured one, two, and three weeks before each auction.Each lag uses an independent seven-day window, while early missing values are median-imputed.
  • Temporal lags:: Elevation, grade, and tier encode the physical and commercial hierarchy of each price observation, while fx_usd captures the LKR/USD exchange rate.These features allow models to distinguish price behaviour across quality levels and catalogues.
  • Temporal lags:: Low Grown teas have the highest average prices, followed by High Grown, Dust, and Off-Grade catalogues.The ordering demonstrates strong price segmentation across catalogues.
  • Temporal lags:: Low Grown contributes 5,384 usable records, while Off-Grade has 15.95% missing price values and uneven categorical completeness across catalogues.These differences can bias aggregate comparisons and make catalogue-specific modelling more stable.
  • Temporal lags:: Because auction prices are not assumed to be stationary, the analysis begins with diagnostic tests before applying regression-based causality methods.This ordering addresses the risk that non-stationary series inflate type-I error.

1) Stationarity:

The analysis checks stationarity before modelling and combines differencing, Granger testing, and tree-based forecasting for catalogue-level weekly prices.

  • Stationarity:: Low Grown prices are non-stationary in levels (p=0.1087) and require first-differencing, whereas High Grown, Off-Grade, and Dust are stationary at levels (p < 0.05).Precipitation, temperature, and sunshine duration are generally stationary across catalogues.
  • Stationarity:: Granger tests evaluate each catalogue, weather variable, and one-to-three-week lag combination after differencing.The resulting claims concern predictive causality in the Granger sense rather than structural causation.
  • Stationarity:: Four tree-based ensembles are compared using unified pooled models and independently trained models for High Grown, Low Grown, Off-Grade, and Dust.The algorithms are Random Forest, Gradient Boosting, XGBoost, and LightGBM.
  • Stationarity:: The forecasting target is price_next_week, created by shifting price_mid_lkr forward one auction week within each catalogue, grade, and tier stream.The horizon reflects pre-auction price expectations.
  • Stationarity:: Models use 5-fold TimeSeriesSplit cross-validation, grid search over eight configurations per model, median imputation pipelines, and raw-LKR predictions.The validation design respects auction chronology and prevents data leakage.

V. RESULTS AND ANALYSIS

Weather-price links are concentrated in particular catalogues and lags, while LightGBM is generally strongest among catalogue-specific models.

  • V. RESULTS AND ANALYSIS: 11 statistically significant weather-price relationships were identified across one-to-three-week Granger lags.Figure 2 summarizes their relative strength using darker heatmap colors for stronger causal influence.
  • V. RESULTS AND ANALYSIS: F=5.275, p=0.0238 for Low Grown precipitation at Lag 1, with significant effects also at Lag 2 (F=4.414, p=0.0147) and Lag 3 (F=3.222, p=0.0263).The relationships concern precipitation in Galle and Matara.
  • V. RESULTS AND ANALYSIS: F=6.097, p=0.0152 for Off-Grade temperature at Lag 1, the strongest single relationship observed.The relationship concerns Uva and Udapussellawa.
  • V. RESULTS AND ANALYSIS: Sunshine and precipitation affect Low Grown prices at Lags 1, 2, and 3, while High Grown shows no significant short-term weather causality.Off-Grade and Dust show strong temperature relationships at three-week lags.
  • V. RESULTS AND ANALYSIS: LightGBM is best for High Grown, Low Grown, and Off-Grade, while Random Forest is best for Dust.The best algorithm therefore varies by catalogue.
  • V. RESULTS AND ANALYSIS: LightGBM is described as the most reliable overall model, especially for Low Grown, where it captures strong and stable patterns while modelling nonlinear price movements.It remains competitive in High Grown, where the signal is weaker.

2) Catalogue-Specific Models:

Catalogue-specific evaluation shows heterogeneous best-performing algorithms, whereas the unified pooled models achieve high accuracy with LightGBM ranking first.

  • Unified Pooled Model:: RMSE = 137.95 LKR for the strongest unified pooled model, LightGBM.The remaining algorithms follow within a narrow performance band.
  • Unified Pooled Model:: R2 > 0.948 for all unified pooled algorithms.The pooled feature set produces strong predictive performance across models.

VI. KEY FINDINGS

Cyclone Ditwah produced a temporary upward movement in Low Grown and High Grown prices, with High Grown showing the strongest short-term spike. The study also contributes a cleaned, public dataset covering 105 weekly sales and 12,233 records.

  • Short-Term Market Response to Extreme Weather: Cyclone Ditwah struck Sri Lanka’s eastern coast on 28 November 2025, bringing flooding and landslides that affected tea-growing areas.
  • Short-Term Market Response to Extreme Weather: High Grown teas displayed the strongest short-term price spike during the cyclone impact window.
  • Short-Term Market Response to Extreme Weather: Both Low Grown and High Grown mid-prices rose toward a mid-window peak before softening, indicating temporary event-period price pressure.
  • Clean and Reproducible Tea Auction Dataset: The cleaned dataset integrates sale-level, price-level, and weather variables through sale_id and rule-based category-to-region mapping.
  • Clean and Reproducible Tea Auction Dataset: The study makes the dataset and code publicly available to support reproducible analysis of Sri Lankan tea auction data.
  • Clean and Reproducible Tea Auction Dataset: 12,233 records with 26 features across 105 sales from November 2023 to March 2026 comprise the final reusable dataset.

VII. DISCUSSION

Weather effects differ across tea catalogues, while catalogue-specific models better reflect these heterogeneous price behaviors. LightGBM performs best for three catalogues, and practical forecasts should retain lagged weather features where supported.

  • Weather Impact on Tea Prices: Low Grown prices respond significantly to precipitation at lags of 1, 2, and 3 weeks, strongest at lag 1 with F=5.275 and p=0.024.The weather signal is largely absorbed into quality and grade rankings by auction day.
  • Weather Impact on Tea Prices: High Grown prices show a visible but transitory response to extreme weather despite no statistically significant lag-based Granger relationships.The adjustment is described as rapidly incorporated into market expectations rather than persisting across longer lag structures.
  • Weather Impact on Tea Prices: Lagged weather features at 1–3 weeks should be included for Low Grown and Off-Grade models but may be omitted for High Grown with minimal accuracy loss.
  • Rationale for Catalogue-Specific Modeling: Unified-model accuracy can be misleading because high-volume, high-variance Low Grown observations may mask weaker performance in other catalogues.
  • Rationale for Catalogue-Specific Modeling: Catalogue-specific LightGBM models perform best for High Grown, Low Grown, and Off-Grade, while Random Forest performs best for Dust.
  • Rationale for Catalogue-Specific Modeling: Selecting models separately by catalogue is presented as essential for practical accuracy and robust trading and policy predictions.

C. Limits of One-Week-Ahead Forecasting.

One-week-ahead forecasting is constrained by auction-day variation that pre-auction features do not capture. The study therefore identifies catalogue heterogeneity and real-time auction signals as important boundaries for forecast improvement.

  • Limits of One-Week-Ahead Forecasting: High Grown and Dust prices have substantial week-ahead variation driven by buyer participation, bidding behavior, and blend-maker decisions absent from pre-auction data.
  • Limits of One-Week-Ahead Forecasting: High Grown and Dust forecasting may require auction-day signals such as buyer activity or real-time bidding data rather than further pre-auction feature refinement.
  • Conclusion: The CTA price series comprises distinct catalogues with varying sensitivities to external drivers rather than one monolithic market series.
  • Conclusion: Off-Grade and Low Grown are identified as the most weather-sensitive catalogues, while High Grown responds briefly to extreme weather shocks.
  • Conclusion: Catalogue-specific models select LightGBM for High Grown, Low Grown, and Off-Grade, and Random Forest for Dust.
  • Conclusion: 137.95 RMSE and 0.9515 R² are reported for unified LightGBM forecasting, which ranks first in pooled accuracy.
Loading 2608.24894v1…