Predictive Analysis of Feed Conversion in Poultry Farming

Context

Feed conversion (kg of feed per kg of chicken produced) is the indicator that defines a poultry integrator's margin. The dataset had 5,813 batches and 40 variables covering the entire production cycle — housing, nutrition, genetics, weekly mortality, slaughter. The business question: can a batch's efficiency be predicted, and what affects it most?

Technical decisions

Choose k by generalisation, not by the best test score. The k with the highest test R² was clearly overfitting (k=5: train R² 0.48 against 0.24 on test). The criterion adopted was the smallest k with ΔR² ≤ 0.05, confirmed by 5-fold cross-validation — it landed on k=25. It is a choice that worsens the number in the table and improves the model.

Fit the preprocessing on the training set only. StandardScaler with fit exclusively on the training set. Without that, information from the test set leaks into the model and the metric lies.

Translate the error into the language of whoever pays. An MAE of 0.0264 means nothing on its own. Applied to a typical batch of 92,000 kg of chicken, it becomes ~2,400 kg of feed — 1.5% of consumption. That is the number the integrator understands.

Result

The linear model (SGD) beat KNN by a wide margin: R² of 0.56 against 0.24, and with less overfitting. The result suggests a predominantly linear and additive relationship between the production variables and feed conversion — a scenario where searching for neighbours in a 99-dimensional space loses effectiveness.

The business insight was worth more than the model: late mortality is the single largest driver of feed conversion (r = 0.50 for cumulative mortality up to day 49). And the correlation grows over the cycle — a bird that dies in week 7 has already eaten almost all of its feed without delivering weight at slaughter. The practical recommendation was to concentrate biosecurity on weeks 5 to 7, a bigger lever than adjustments to nutrition or genetics.

Learnings

An R² of 0.56 is a modest ceiling — weather, feed batch quality and daily handling were outside the dataset. Tree-based models would probably capture interactions that linear regression cannot see, but they were outside the scope of the course. And there is a real-world limitation: some of the variables are only known at the end of the cycle, which restricts predictive use before slaughter. A next version would use only what is known on housing day.