What changed?
Previously, our simulator modeled yield by tracking individual fruits digitally. While precise, this method may infer a more explicit fruit set or development time trend than for the greenhouse population. As a result we often saw a "zig-zag" volatility in forecasts.
We have replaced this with a continuous modelling approach which decreases volatility.
From discrete to continuous: Instead of simulating single fruits, our simulator now applies a distribution in set / development time to fruits.
Impact
While we release this all harvest forecasts will suddenly change. You will see a change in what we predict for the next couple of weeks.
Higher accuracy: We have observed a ~4% percentage error decrease and 10% reduction in outliers across all forecasts. Any prediction larger than 20% is considered an outlier.
Increased stability: The forecast is less prone to sudden jumps caused by noisy data. This leads to a smoother, more reliable trend line closer to what actually happens in the greenhouse.