AgriPriceBD: A New Dataset Challenges Commodity Forecasting in Bangladesh

Forecasting agricultural commodity prices is no simple task, especially in developing regions where data scarcity is a real hurdle. Enter AgriPriceBD, a new dataset aiming to transform how we predict prices for key Bangladeshi commodities. Spanning from July 2020 to June 2025, this set captures daily mid-prices for garlic, chickpea, green chilli, cucumber, and sweet pumpkin. It's extracted through an LLM-assisted pipeline from government reports.

Why AgriPriceBD?

This dataset's emergence is significant. South Asia needs machine-learning-ready data for effective commodity price predictions. Without it, the economic stability of smallholders hangs in the balance. Yet, why should we assume AgriPriceBD is the silver bullet? The dataset's ability to feed various forecasting models could be a major shift for agricultural markets in Bangladesh and similar economies, but only if the models perform as promised.

Evaluating Forecasting Models

The paper's key contribution: a comprehensive evaluation of seven forecasting approaches. These encompass both classical models like SARIMA and Prophet, and deep learning architectures like BiLSTM and Transformer derivatives. The assessments were backed by Diebold-Mariano statistical tests.

An unexpected revelation? Naïve persistence is more effective for near-random-walk commodities. It's a fact that challenges the hype around complex models. Time2Vec, once touted as a star in temporal encoding, falls flat. Its integration leads to a catastrophic 146.1% MAE increase for green chilli predictions.

Prophet and Informer: Falling Short

Prophet, another much-discussed tool, systematically fails here. Its smooth decomposition assumptions clash with the discrete nature of price dynamics in these commodities. Informer's performance? Erratic at best, with predictions showing variance up to 50 times the ground truth. This highlights a critical issue: sparse-attention Transformers need more extensive datasets than the currently available small-scale agricultural ones.

Looking Forward

What does this mean for future research? The key finding is that while AgriPriceBD sets the stage, it exposes the limitations of existing models on small datasets. The promise of sophisticated methods like Informer remains unmet without broader data availability.

So, where does that leave us? The dataset, along with all related code and models, is publicly available. It's an open call to the research community to refine and build upon these initial findings. Will it spur the development of models tailored to smallholder needs? That's the real question.