❄️ Sales Prediction Challenge — AC sales in Mohali
Some inputs really drive AC sales — others are redundant, noise, or even a trap that looks great on training days but fails on new ones. Pick the inputs that actually help, shape the layers, and train (max 50 epochs). Your Score = test MSE × (1 + 0.4% per parameter), so the winner gets the lowest error with the smallest network — every extra input or neuron costs you. Everyone shares the same data & test split. Sign in, beat your best, and submit — your teacher sees the leaderboard.
Each dot is a day, with weather drawn from Mohali's real seasonal climate (Chandigarh climate normals) — hot dry summers, humid monsoon, cool winters.
FEATURES
Which inputs actually help? Not all of them do. Click to toggle. 2 selected.
Line blue = positive weight, orange = negative; thickness = strength. Watch them shift as it trains.
OUTPUT
• train • test — perfect fit
🧪 How to win: low error, few neurons
Beware the trap feature
Watch train vs test MSE. A feature that makes train error tiny but test error worse is a trap — it fit noise that won't repeat. Keep only inputs that lower test MSE.
Smaller network wins ties
Your Score = test MSE × (1 + 0.4% × parameters). Every extra input or neuron adds parameters. A tiny model that nearly matches a big one wins. Trim until the fit just starts to suffer.
Only 50 epochs
Everyone gets the same budget. Use a higher learning rate (0.1–0.3) so it converges in time — but too high and the error bounces around.
Read the diagram
Thick blue/orange lines are strong weights. A useless input usually ends up with thin, faint lines — a hint you can drop it.