• Guilherme@lemm.ee
    link
    fedilink
    English
    arrow-up
    2
    ·
    20 days ago

    I browse Windy but don’t rely on any of those 4 weather forecasting models: I take the median of predicted temperatures and rainfalls instead. Also, I predict rain only if the median exceeds 1mm, and if it’s below that threshold but at least 3 models predict (some) rainfall I predict drizzle. Which is the same approach I had at my previous job, using data of doubtful quality to adjust Holt-Winters and Box-Jenkins models in order to forecast drug sales for Big Pharma.

    Kaggle by the way began to demand users engaged on modelling competitions to make PDFs explaining their methodologies after learning some cheaters would just combine results from other competitors.

    P.S. - Don’t average results from different models unless you are really, really sure of what you’re doing. Many times the models take turns on which one will output garbage, and you don’t want garbage contaminating your average. By switching to median you avoid the crap they sometimes spit altogether - not to mention it’s so simple you don’t even need to write numbers on paper or use a pocket calculator.