All projects

Case study · Data Science

Chicken Price Forecasting System

A weekly price forecasting pipeline for 15 Indonesian provinces, built so that production could be planned before market prices were known.

Context
Poultry industry, production forecasting; open-source reimplementation
Stack
  • XGBoost
  • SHAP
  • Streamlit
  • Python
  • SAP
  • scikit-learn
  • Time Series
Annual savingsReported business impact
$3M+
MAPE on held-out weeksPublic synthetic benchmark
3.2%
Provinces forecast weekly
15

The problem

Chicken prices in Indonesia move sharply with seasons, holidays, feed costs and supply shocks, and they differ by province.

Production decisions have to be made weeks ahead, so the business needed a forecast per province, together with the reasons behind each prediction.

Data

Weekly prices for 15 provinces, with feed costs, supply, weather severity, holidays and regional logistics costs.

The public repository runs on a synthetic dataset of 104 weeks × 15 provinces that reproduces these patterns, in place of the original business data.

Approach

  1. 01

    Data extraction

    Weekly price and cost data pulled from SAP into a single modelling table.

  2. 02

    Feature engineering

    23 features covering price and feed lags, rolling statistics, seasonality and holiday or weather interactions, defined once and shared by training and inference.

  3. 03

    Modelling

    XGBoost regression benchmarked against a linear baseline on a time-based 85/15 split.

  4. 04

    Explainability

    SHAP values for every prediction, so planners can see which drivers moved the forecast.

  5. 05

    Delivery

    A Streamlit dashboard with province charts and alerts, refreshed by a weekly scheduled run that flags degradation when RMSE exceeds 1,200 IDR/kg.

Impact

When the model forecast a price decline, the business redirected fresh chicken to its frozen product lines instead of selling into a falling market.

Next steps

  • Validate with rolling-origin backtests and keep every price-derived feature strictly lagged.
  • Benchmark against seasonal and statistical baselines and report accuracy per province.
  • Add prediction intervals so planners can see the uncertainty around each forecast.

Figures marked as the public benchmark come from the open-source version, which runs on synthetic data.