# Auction Price Prediction

> Machine learning prediction of auction price from spectral profile. Development paused pending a larger, validated calibration dataset.

- Canonical URL: https://spectrify.ai/research-pipeline/auction-price-prediction

- Sector: Sector 01 · Food & Agriculture
- Status: Paused
- Tags: ML, Tea, Auction, Price Prediction, Gradient Boosting

Auction price prediction from NIR spectral data represents the highest-value potential application in the tea quality stack - the ability to tell a factory, before the auction, what their lot is likely to sell for. Early experiments were promising, but the complexity of auction pricing requires a substantially larger and more temporally diverse dataset than is currently available.

## What We've Established

### Spectral Quality Proxy Correlation

NIR-predicted quality parameters (TPP, moisture, colour score) show statistically significant correlation with Colombo auction lot prices in our initial dataset. Lots with high TPP and optimal moisture show a consistent price premium. However, the predictive accuracy degrades significantly when controlling for non-spectral variables such as estate reputation, seasonal supply, and buyer mandate preferences.

### Non-Spectral Confounders

A critical finding is that auction price is influenced by at least as many market factors as quality factors. Grade classification, estate origin, seasonal scarcity, and individual buyer mandates can account for large price swings that no quality-based spectral model can predict. This limits the application to price-band estimation rather than precise point prediction.

## Technical Approach

The initial approach used gradient boosting regression on a combined feature set: NIR-predicted quality parameters alongside categorical features (grade, estate region, season). Cross-validation showed acceptable performance within a single estate over a short time horizon but poor generalisation across estates and auction sessions. Direct end-to-end regression from raw spectral features to price is also being explored.

## Where We Stand

Development is paused. The core technical problem is a dataset size and diversity issue - not a modelling ceiling. We require a minimum of 1,500–2,000 paired auction records (spectral scan + validated reference data + realised auction price) spanning at least two full seasons before a generalisable model can be trained. Collecting this dataset is a medium-term data partnership objective.

## Next Steps

- Define data collection protocol for systematic spectral-auction price pairing
- Engage auction-house stakeholders for structured data-sharing discussions
- Revisit feasibility once dataset threshold is met
- In the meantime, explore price-band classification as a lower-precision but more immediately viable output

## Related Projects

- [Total Polyphenols (TPP)](https://spectrify.ai/research-pipeline/total-polyphenols-tpp) (Model Ready): Total polyphenol profiling for black tea - the leading proxy for antioxidant quality and a primary auction pricing signal.
- [Moisture Content - Green Tea](https://spectrify.ai/research-pipeline/moisture-content-green-tea) (Model Ready): Real-time MC measurement for green tea. Model trained on TRI-standard data and independently validated in December 2025.
- [Sugar in Tea](https://spectrify.ai/research-pipeline/sugar-in-tea) (Model Ready): Total sugar quantification across black tea grades - a key differentiator for premium blend specification and export documentation.

[Research Pipeline](https://spectrify.ai/research-pipeline)