Neural networks, deep learning, NLP · insight-0446

Compare a sequence model of booking curves against lag features in a GBM

Situation
You want an LSTM or Transformer to forecast final bookings from the sequence of booking-curve snapshots per flight.
Action
Build snapshot features at fixed days-to-departure (cumulative bookings, pickup since last snapshot, same flight last year) and train a GBM under the same rolling-origin backtest and metric. Adopt the sequence model only if it wins per horizon.
Reason
On standard long-horizon benchmarks, simple linear models outperformed several Transformer forecasters, showing that architecture complexity alone does not guarantee better forecasts.
Boundary
Forecasting from fixed-length, regularly spaced histories, such as booking curves or sensor windows.

Symptom or risk

Sequence models need more data and tuning than lag-feature models; without a baseline, a complex model ships with no evidence it helps.

Irregular event sequences or very large multi-series datasets can favour sequence models.

Lag features need careful point-in-time construction (Day 24); sequence models need careful masking.

Using snapshots recorded after the forecast cutoff in either model.

Zeng et al. found simple linear models (DLinear/NLinear) outperformed several Transformer-based long-term forecasting models on standard benchmarks.
Are Transformers Effective for Time Series Forecasting?