← Back to Dashboard

Triple-Model AI Pipeline

Our forecasting engine combines three fundamentally different AI architectures to produce consensus predictions, reducing the risk of any single model's blind spots.

1
Data Collection

2,869+ records of daily cardamom market data, weather, and supply chain metrics

2
Parallel Inference

Three models process data simultaneously with different perspectives

3
Consensus

Results are combined using weighted voting based on model confidence

4
Decision

Actionable recommendations generated using a multi-factor decision matrix

The Models

Time Series Foundation

🕐 Google TimesFM

A foundation model for time-series forecasting pre-trained on billions of data points. It excels at capturing long-horizon trends, seasonal patterns, and regime changes in the cardamom market. Serves as the price-direction backbone of our pipeline.

Deep Learning

🧠 LSTM-CNN 34-Layer

A hybrid architecture combining Long Short-Term Memory (LSTM) networks for sequential pattern recognition with Convolutional Neural Networks (CNN) for spatial feature extraction. With 34 layers, this model captures complex non-linear relationships between weather, supply, and price volatility.

Gradient Boosting

📊 XGBoost v17

An ensemble gradient-boosted decision tree model trained specifically on tabular cardamom market features. It excels at binary buy/sell classification using features like auction quantities, weather lags, and price momentum. Version 17 includes flood and lockdown indicators.

Decision Matrix

Recommendations are generated using a multi-signal approach:

Data Sources

Historical data sourced from the Indian Spice Board auction records (Bodinayakanur market), supplemented with IMD weather data, soil moisture indices, and supply chain disruption flags. 21 engineered features including lagged prices, rolling averages, EMAs, and binary event indicators.

← Return to Dashboard