Cellular Predictions on the Move: What about Data?

The paper's actual contribution is narrower than the summary suggests: it's not that data matters (practitioners already know this), but rather that cellular load forecasting specifically benefits from behavioral and demographic signals that traditional base-station logs omit. The claim is domain-specific, not a general indictment of model-first approaches.

This work sits alongside the Black-Box Assisted Regression paper from earlier today, which quantifies exactly when labeled data collection outpaces foundation model priors in specialized tasks. Both papers are converging on a production reality: practitioners deploying models on narrow infrastructure problems face a concrete tradeoff between collecting richer features and tuning architecture. The cellular forecasting work provides a concrete case study of that tradeoff in action, showing that behavioral signals (population movement, event calendars) beat architectural innovation when base-station telemetry alone is insufficient.

If the authors release a public benchmark comparing their data-enriched baseline against standard LSTM or Transformer baselines on the same cellular dataset, and the gap persists when both models have equal hyperparameter tuning budget, that confirms the finding is about feature engineering rather than model selection. If the gap shrinks when the baseline model gets access to the same behavioral signals, the story collapses to 'more data helps,' which is not novel.