colombo-tea-auction-prices/cta-price-prediction

Dataset Card for Weather-Driven Sri Lankan Tea Market Catalogues

Dataset Description

This dataset contains 12,233 structured records extracted from 105 weekly Forbes & Walker Tea Brokers PDF reports spanning from November 2023 to March 2026. It represents the first machine-readable, comprehensive archive of the Colombo Tea Auction (CTA) prices paired with localized, region-specific lagged weather variables from the Open-Meteo historical archive.

The dataset is designed to evaluate the short-term and long-term predictive associations between highly localized weather anomalies and agricultural commodity auction price dynamics across heterogeneous commercial market catalogues.

• Repository: https://github.com/hesandism/data-analysis-for-tea-industry
• Pre-print: http://arxiv.org/abs/2608.24894
• Paper: Forecasting Weather-Driven Price Dynamics Across Sri Lankan Tea Market Catalogues (MERCon 2026)

Dataset Structure

Data Overview

• Total Rows: 12,233 rows across 105 weekly auction sales.
• Primary Key: sale_id (used for matching market and weather records).

Target Variables

• price_mid_lkr: Derived as (price_lo_lkr + price_hi_lkr) / 2 (with a fallback to price_lo_lkr if only a lower bound is reported).
• price_next_week: The target variable for 1-week-ahead forecasting, shifted forward by one auction week within each distinct product stream.

Key Catalogues & Statistics

The dataset is split across 4 distinct commercial catalogues reflecting sorting qualities and elevation metrics:

Catalogue / Group	Total Rows	Mean Price (LKR)	Median Price (LKR)	SD (LKR)	Missing Target (%)	Primary Weather Region Alignment
Low Grown	5,384	1,578.7	1,400.0	749.5	0.0%	Low Grown/Matara-Galle (6.2500°N, 80.3000°E)
High Grown	2,124	1,135.6	1,170.0	261.5	0.0%	Western High Grown/Maskeliya (6.9271°N, 80.5350°E)
Dust	2,205	984.4	985.0	207.3	1.5%	Sourced globally; Uses Low Grown (Primary), Western High (Secondary)
Off-Grade	2,520	792.2	730.0	169.3	16.0%	Sourced globally; Uses Low Grown (Primary), Western High (Secondary)

Features Included (Total: 26 Features)

1. Market Metrics: Average prices (LKR, USD), volume sold (kg), crop intake directions, and USD/LKR consolidated exchange rates (fx_usd).
2. Text-Parsed Weather: Textual keyword match arrays, and calculated composite avg_weather_severity scores extracted via natural language pipelines from broker report text sections.
3. Meteorological Lags (Open-Meteo API): 7-day window aggregations for Total Precipitation (mm), Mean Temperature (°C), and Sunshine Duration (seconds) lagged deliberately at 1, 2, and 3 weeks prior to the auction (auction_date - 7xlag) to isolate biological and logistical operational delays.

Methodology & Pipeline

The data was generated using an end-to-end 5-stage pipeline:

1. Data Collection: 105 semi-structured PDF documents processed using pdfplumber to bypass embedded conversational texts and isolate tabular price grids.
2. Feature Engineering: Calculation of empirical mid-prices, structural encoding (tier, grade, elevation indices), and temporal lag extractions.
3. Diagnostics: Stationarity checking via Augmented Dickey-Fuller (ADF) showing Low Grown profiles require first-differencing ($p=0.1087$) while other tiers hold stationarity in baseline levels ($p<0.05$).
4. Granger Causality Evaluation: Multi-lag sensitivity checking spanning 1–8 weeks under False Discovery Rate adjustments (Benjamini-Hochberg).
5. Machine Learning Benchmarking: Optimization blocks comparing baseline estimators (Naive Persistence, AR(1)) with ensemble deciders (Random Forest, Gradient Boosting, XGBoost, LightGBM).

Key Research Insights

• Market-Dominant Signals: In one-week-ahead auction windows, structural market factors and localized pricing persistence dictate the vast majority of directional trend configurations. For instance, Naive Persistence was the most optimal forecasting method for the Low Grown catalogue ($RMSE = 79.76$, $R^2 = 0.9896$).
• Robust Weather Causality: Extended lag checking isolated exactly one robust, FDR-significant connection surviving conservative Benjamini-Hochberg correction ($q < 0.05$): cumulative ambient temperature variations over a 4-week history Granger-predict price shifts uniquely for the Dust market catalogue ($F=5.843, p=0.00031, q=0.0257$).
• Extreme Shocks: The repository details an event-window trace recording the sudden macroeconomic impact of Cyclone Ditwah (Nov 2025–Jan 2026), exposing a steep, volatile, yet highly transient short-term price correction premium concentrated heavily inside the High Grown sector.

Usage & Replication

import pandas as pd

# Load dataset (Replace with actual loading method when hosted)
df = pd.read_csv("colombo_tea_auction_master.csv")

# Filter for the High Grown catalogue
high_grown_df = df[df['catalogue'] == 'High Grown']

# Baseline first-order Autoregression setup
high_grown_df['price_lag_1'] = high_grown_df.groupby('product_stream')['price_mid_lkr'].shift(1)

Citation

If you use this dataset or repository in your research, please cite:

@article{cta2026forecasting,
  title={Forecasting Weather-Driven Price Dynamics Across Sri Lankan Tea Market Catalogues},
  author={Mallawarachchi, Hesandi and Madurapperumage, Senilka and Kulathunge, Nadil and Angeesa, Thilokya and Gunaweera, Nethsith and Weerasekara, Sandeepa and Narasinghe, Patalee and de Silva, Nisansa and Wickramanayake, Sandareka},
  journal={arXiv preprint arXiv:2608.24894},
  year={2026},
  doi={10.48550/arXiv.2608.24894}
}