💰 quantitative finance

Kalimati Vegetable Price Index Forecasting with a Momentum Corrected Online Stacking Ensemble

This study introduces the Kalimati Vegetable Price Index (KVPI) and a Momentum-Corrected Online Stacking Ensemble model that achieves highly accurate 90-day price forecasts (0.68% MAPE) for agricultural commodities in Nepal, offering a robust tool for policymakers to enhance food security amidst market volatility.

Original authors: Sahaj Raj Malla

Published 2026-06-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sahaj Raj Malla

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather in a chaotic, mountainous region where the wind changes direction every hour, and sudden storms happen because of local festivals. Now, imagine trying to do this not for the weather, but for the price of vegetables in Kathmandu, Nepal. That is exactly what this paper attempts to solve.

Here is the story of the research, broken down into simple concepts and analogies.

1. The Problem: The "Noisy" Vegetable Market

In Nepal, the price of vegetables is incredibly unstable. It jumps up and down wildly due to:

Monsoon rains washing out roads.
Festivals (like Dashain or Tihar) where everyone suddenly wants to buy more food.
Supply chain hiccups.

Trying to predict the price of just one vegetable (like tomatoes) is like trying to hear a single person whispering in a crowded, screaming stadium. The signal is too noisy. The data is messy, missing pieces, and full of "static."

2. The Solution: Creating a "Weather Map" (The KVPI)

Instead of listening to one whispering vegetable, the researchers decided to listen to the whole crowd. They created a new tool called the Kalimati Vegetable Price Index (KVPI).

The Analogy: Imagine you have 135 different vegetables. Some are very volatile (like onions that change price daily), and some are steady (like potatoes). The researchers didn't just average them out. They built a "smart average."
How it works: They gave less "volume" to the vegetables that scream the loudest (high volatility) and more "volume" to the steady ones. This created a smooth, stable "macro-signal" that represents the overall health of the vegetable market, filtering out the individual noise of single crops.

3. The Toolkit: 64 Clues (Features)

To predict the future, the researchers didn't just look at past prices. They gathered 64 different "clues" or hints, ensuring they didn't cheat by looking at the answer key (a problem called "look-ahead bias").

The Clues included:
- History: What happened 1 day ago? 7 days ago? 30 days ago?
- Rolling Stats: What was the average price over the last week?
- The Festival Calendar: They specifically marked dates for big Nepali festivals. They knew that prices often spike before a festival (people stocking up) and drop after (the rush is over).
- Time: Is it a Saturday? Is it a holiday?

4. The Contest: 14 Different Forecasters

The researchers set up a competition with 14 different "forecasters" (computer models) to see who could predict the vegetable prices best. They tested them on short-term (1 week), medium-term (1 month), and long-term (3 months) predictions.

The Old Guard (Statistical Models): These were like experienced accountants using simple math rules (ARIMA). They were good at steady trends but got confused when the market got crazy or when festivals hit.
The Deep Learners (AI/Transformers): These were like super-smart robots trained on massive datasets. However, because the vegetable data was "noisy" and not huge, these robots got confused and started "hallucinating" (overfitting), making terrible predictions.
The Tree-Ensembles: These were like a committee of decision-makers (XGBoost, Random Forest). They were very good at handling messy data and complex rules.

5. The Winner: The "Momentum-Corrected" Team

The champion wasn't a single model, but a teamwork approach called the Momentum-Corrected Online Stacking Ensemble.

How it works (The Analogy): Imagine a relay race team.
- The Runners: The team combines the best "Tree" models (good at spotting sudden spikes) and the best "Recurrent" models (good at spotting long-term trends).
- The Coach (Momentum Correction): This is the secret sauce. Usually, prediction teams are slow to react when the market suddenly changes direction (like a sudden price spike before a festival).
- The Trick: This coach watches the errors the team made in the last few days. If the team keeps underestimating the price (the error slope is going up), the coach shouts, "Add momentum!" and pushes the prediction up before the next spike happens. It's like a surfer adjusting their board based on the wave's current speed, not just where the wave was a minute ago.

6. The Results: A Clear Victory

The results were impressive, especially for long-term predictions (90 days):

Accuracy: The winning model was incredibly precise, with an error rate of less than 1% (0.68%).
Comparison: The "Old Guard" statistical models and the fancy "Deep Learning" robots failed miserably at the 90-day mark, often predicting the opposite of what happened. The "Momentum-Corrected" team, however, got it right.
Why it worked: It combined the ability to handle messy data (like the tree models) with the ability to adapt quickly to changing trends (the momentum correction).

Summary

This paper says: "If you want to predict vegetable prices in a chaotic, festival-driven market like Nepal, don't rely on simple math or over-complicated AI. Instead, build a stable index of all vegetables, feed it lots of cultural and time-based clues, and use a smart team of models that can 'feel' the momentum of the market to correct their own mistakes in real-time."

The researchers have made their code and data public, so anyone can use this "smart weather map" to help farmers and policymakers in Nepal plan better and avoid food shortages.

Technical Summary: Kalimati Vegetable Price Index Forecasting with a Momentum Corrected Online Stacking Ensemble

Problem Statement
Forecasting agricultural commodity prices in emerging economies like Nepal is hindered by high volatility, supply chain fragility, monsoon dependence, and culturally driven demand surges (e.g., festivals). Existing literature often focuses on individual commodities, a "atomistic" approach that is vulnerable to commodity-specific shocks, data sparsity, and idiosyncratic noise, limiting model generalizability. Furthermore, standard modeling paradigms face distinct challenges in this context: classical statistical models (e.g., ARIMA) struggle with non-linear interactions and high-dimensional exogenous factors, while state-of-the-art deep learning models (e.g., Transformers) often suffer from overfitting and poor generalization on the moderately sized, noisy time series typical of developing markets. No single paradigm has yet proven sufficient for robust, multi-horizon forecasting in these volatile environments.

Methodology
The study proposes a comprehensive framework centered on a new macro-level index and an adaptive ensemble learning strategy.

Data Construction (KVPI):
- Source: Daily wholesale price records from the Kalimati Fruits and Vegetables Market (Kathmandu) covering June 2013 to September 2023 (280,939 records, 136 commodities).
- Index Creation: To mitigate noise, the authors constructed the Kalimati Vegetable Price Index (KVPI). This is an inverse-volatility weighted composite index aggregating 135 commodities (those with >365 days of data).
- Weighting: Commodities are weighted inversely to their coefficient of variation ( $CV_c$ ) to reduce the influence of erratic items. The index spans 3,757 days with a range of [90.1, 219.9].
- Preprocessing: Rigorous cleaning included outlier capping (winsorization) based on training-set IQR bounds, missing value imputation (forward-fill and bidirectional interpolation), and strict separation of training (2013–2022) and testing (2022–2023) sets to prevent data leakage.
Feature Engineering:
- A high-dimensional matrix of 64 causally valid predictors was engineered.
- Components: Autoregressive lags (1–30 days), rolling statistics (mean, std, min, max, median), exponentially weighted moving averages, differencing, calendar features (including Nepal-specific weekend indicators), and festival dummy variables with lead/lag windows for major events (Dashain, Tihar, etc.).
- Constraint: All features were shifted backward by one day to strictly eliminate look-ahead bias. Non-causal global smoothers were excluded.
Modeling and Evaluation:
- Benchmarks: 14 architectures were evaluated across four horizons ( $h \in \{7, 14, 30, 90\}$ days): Naïve baselines, Statistical models (Auto-ARIMA, SARIMA), Tree-based ensembles (Random Forest, ExtraTrees, HistGB, XGBoost), Deep Learning (LSTM, GRU), Hybrid models (ARIMA-LSTM, ARIMA-HistGB), and State-of-the-Art Transformers (PatchTST, NBEATSx).
- Proposed Solution: A Momentum-Corrected Online Stacking Ensemble.
  - Mechanism: This meta-learner uses a causal online mechanism. At each step $t$ , it computes a rolling residual derivative (slope of recent forecast errors) over a window $w$ .
  - Correction: If a systematic bias is detected ( $\beta > 0$ ), a momentum penalty is applied to the base blend prediction.
  - Optimization: A combinatorial search identified an optimal blend of ExtraTrees (for high-frequency volatility capture) and GRU (for trend recognition), with dynamic weighting that effectively zeroed out extraneous models.

Key Results
The proposed Momentum-Corrected Online Stacking Ensemble demonstrated superior performance across all horizons, particularly in the long term.

Long-Term Horizon (90-day):
- RMSE: 1.771 (lowest among all models).
- MAPE: 0.684%.
- $R^2$ : 0.845 (explaining 84.5% of variance).
- Comparison: This significantly outperformed the best individual model, XGBoost ( $R^2 = 0.824$ ), and vastly exceeded statistical baselines and Transformers, which often produced negative $R^2$ values at this horizon (e.g., NBEATSx $R^2 = -21.371$ ).
Short-to-Medium Term: The ensemble maintained the lowest RMSE at $h=7$ (2.330), $h=14$ (2.237), and $h=30$ (1.828).
Model Hierarchy:
- Tree-based ensembles proved the most robust individual competitors.
- Deep Learning (LSTM/GRU): With rigorous hyperparameter tuning (100 trials) and full feature access, these models achieved competitive results ( $R^2 \approx 0.77$ ), challenging the notion that they strictly require massive datasets.
- Transformers (PatchTST, NBEATSx): Struggled with the noisy, culturally driven data, exhibiting severe performance degradation.
- Statistical Models: Showed severe degradation over longer horizons due to linearity assumptions.

Significance and Claims
The paper claims that the Momentum-Corrected Online Stacking Ensemble offers a practical, reliable tool for anticipating price movements in volatile emerging markets. Its primary significance lies in:

Stability: By aggregating 135 commodities into the KVPI, the study creates a stable macro-level signal that reduces noise inherent in individual crop modeling.
Adaptability: The momentum correction mechanism dynamically adjusts for systematic bias during structural breaks and cultural demand shocks (e.g., festival spikes), addressing the "lagging" behavior common in static ensembles.
Methodological Insight: The study demonstrates that for moderately sized, noisy economic time series, a carefully tuned ensemble of tree-based and recurrent models outperforms complex, parameter-heavy Transformer architectures and classical statistical methods.
Policy Utility: The open-source pipeline provides policymakers and supply chain actors in Nepal and similar markets with a reproducible foundation to enhance market transparency, support evidence-based decisions, and strengthen food security.

The authors note limitations, including the reliance on historical price structures without explicit exogenous drivers like weather or fuel costs, and the need for longer validation windows across diverse economic regimes. However, the framework is presented as a robust foundation for future extensions.