📈 LSTM Airline Passenger Forecasting Lab

About this simulation

This scene mirrors the classic "AirPassengers" LSTM exercise: monthly international airline passenger totals from 1949–1960, laid out as a bar chart in 3D. A translucent window sweeps across the bars representing the lookback window an LSTM consumes at each step, while a cluster of glowing rods above it stands in for the network's hidden state reacting to the sequence. A blue tube traces the network's one-step-ahead prediction against the grey actual bars.

How it works

• Lookback window — how many past months the model conditions on. Short windows react quickly but look jagged; long windows smooth and lag behind sharp turns, just like a real recurrent network trained with short vs. long input sequences.
• Hidden units — the network's representational capacity. With few units the prediction under-fits and hugs a smoothed trend; with more units it captures more of the seasonal wiggle and tracks the actual curve more closely.
• Forecast horizon — beyond the last observed month, the model has no ground truth to lean on, so its own predictions feed back in as input. The amber cone shows how uncertainty compounds the further out it forecasts, shrinking only with more hidden capacity.

Controls

• Lookback window — number of months of context fed to the model (2–24).
• Hidden units — LSTM hidden-state size (2–64).
• Playback speed — how fast the reading cursor sweeps through the series.
• Extend forecast — reveal 12 months of pure forecast beyond the training data.
• ↺ Restart sweep — resets the cursor to January 1949. Drag to rotate, scroll to zoom.

Did you know?

The real AirPassengers dataset (Box & Jenkins, 1976) is a textbook example of multiplicative seasonality with trend — exactly the kind of nonstationary pattern that made simple ARIMA models popular for decades before recurrent networks like LSTMs could learn the same structure directly from raw sequences.