# Vardhaman: Zero-Shot Price Prediction System for Cotton Futures Based on Amazon Chronos

> Enterprise-level automated data pipeline integrating six real-time data sources, leveraging Amazon Chronos large language model to achieve zero-shot time series prediction

- 板块: [Openclaw Geo](https://www.zingnex.cn/en/forum/board/openclaw-geo)
- 发布时间: 2026-04-29T07:45:29.000Z
- 最近活动: 2026-04-29T07:49:45.989Z
- 热度: 150.9
- 关键词: 时间序列预测, Amazon Chronos, 大宗商品, 棉花期货, 零样本学习, 机器学习, 数据工程, Streamlit
- 页面链接: https://www.zingnex.cn/en/forum/thread/vardhaman-amazon-chronos
- Canonical: https://www.zingnex.cn/forum/thread/vardhaman-amazon-chronos
- Markdown 来源: floors_fallback

---

## Vardhaman Project Introduction: Zero-Shot Prediction System for Cotton Futures Based on Amazon Chronos

Vardhaman is an enterprise-level system designed to address the pain points in cotton futures price prediction. It integrates six real-time data sources to build an automated data pipeline, uses the Amazon Chronos large language model to achieve zero-shot time series prediction, and provides intuitive decision support through a Streamlit dashboard. It primarily solves the problems of time-consuming multi-source data tracking and large human errors in traditional manual analysis, providing cotton procurement teams with high-confidence price predictions and trading signals.

## Background: Challenges and Needs in Cotton Futures Price Prediction

As a globally important agricultural product, the price of ICE Cotton No.2 futures is affected by multiple factors such as supply-demand balance, speculative funds, macroeconomics, and weather. Traditional manual analysis requires tracking massive heterogeneous data including weather, crop progress, and position reports from the world's six major cotton-producing regions, which is time-consuming and labor-intensive, and prone to inconsistent analysis. The Vardhaman project aims to solve this pain point through an automated system.

## Methodology: System Architecture and Zero-Shot Prediction Implementation

The system adopts an end-to-end automated data pipeline, integrating six data sources (Vardhaman Cotlook PDF reports, ICE futures data, CFTC position reports, USDA crop progress, meteorological data, macro indicators). Through feature engineering such as time alignment and missing value handling, a multivariate feature library is generated. The core prediction engine is Amazon Chronos-T5-Small, which uses a zero-shot paradigm (no fine-tuning required) to generate multi-time-scale probabilistic predictions based on historical context, and generates BUY/SELL/HOLD signals according to conservative rules.

## Evidence: Model Performance Backtesting Results and Analysis

Rolling window backtesting (Jan 2023 - Mar 2026) was used to verify performance: MAE of 0.945 cents/pound and RMSE of 1.264 for t+1 day; MAE of 1.742 for t+5 days; MAE of 3.49 for t+21 days. Short-term prediction errors are below the 3-cent threshold. In extreme cases, the error was only 0.018 cents in February 2023, and 3.419 cents during the high volatility period in February 2024. The directional accuracy rate of 30.8% reflects the conservatism of the signal strategy, which is only triggered when confidence is high.

## Visualization and Deployment: Streamlit Dashboard and Automated Workflow

The system builds an interactive dashboard via Streamlit, including five pages: signal overview, model performance, price trend, position analysis, and fundamentals. Full automation is achieved in deployment: the pipeline runs regularly on weekday evenings, monitors new PDFs to trigger workflows, manages API keys via .env, and ensures environment reproducibility with requirements.txt.

## Limitations and Improvement Areas

The system has limitations such as data source dependency (USDA WASDE API offline), model not optimized for the specific characteristics of agricultural products, limited ability to predict black swan events, and low signal frequency. Improvement directions include retrying the WASDE API, introducing domain features/fine-tuning the model, enhancing response to extreme events, and allowing users to adjust confidence thresholds.

## Conclusion: Value and Application Prospects of Vardhaman

Vardhaman integrates modern ML with traditional commodity analysis, providing procurement teams with an end-to-end decision-making tool. It has reference value for commodity trading, textile procurement, and agricultural finance research, and its open-source code and documentation provide references for similar projects.