With 120 daily observations from a single ATM, would you ship a recurrent forecaster or a classical model?

## Frame it as a data-budget question The instinct to reach for a sequence model comes from the sequence in the name. The right first question is how much data the model actually gets to learn from, and for a single 120-point daily series the answer is brutal. - **Cycles observed.** With a day-of-week pattern, 120 days is about seventeen repetitions. Everything the model can learn about weekly structure it must learn from seventeen noisy examples of it. - **Training examples.** Window it with a 14-day lookback and a 7-day horizon and you get roughly a hundred windows, each overlapping its neighbours almost completely. The effective sample size is nearer seventeen than a hundred. - **Parameters.** A vanilla recurrent cell with hidden size 32 on a single input channel carries 32 by 32 recurrent weights plus input weights and biases, about 1,100 parameters. A gated cell has four such parameter blocks, so over 4,000 — before the output layer, before a second layer. You are fitting thousands of free parameters to a few hundred target values. No amount of dropout or weight decay closes that gap. Regularisation narrows the hypothesis space; it does not manufacture information. ## What the classical model brings A seasonal statistical model — exponential smoothing with a weekly seasonal component, or a seasonal ARIMA fit — carries on the order of ten parameters. Its structure encodes the assumption that the future resembles a level, a trend and a repeating weekly pattern, which is exactly the prior a short series cannot afford to learn from scratch. It produces prediction intervals essentially for free, it fits in milliseconds so refitting nightly is trivial, and there is no tuning search to overfit. And before either: **seasonal-naive** — this Tuesday equals last Tuesday. On a strongly weekly ATM series that baseline is surprisingly hard to beat, and any model that does not beat it convincingly across several rolling origins should not be deployed. Reporting a neural model that loses to seasonal-naive is a bad day; not having run the baseline at all is worse. ## The evaluation trap With 120 points, whatever you hold out is small. A single 20-point holdout has an error estimate so noisy that the ranking between two models is close to a coin flip, and comparing five architectures on it guarantees the winner is the luckiest, not the best. Two consequences follow. First, prefer models whose selection requires few decisions — a classical family with an order-selection criterion makes far fewer choices than an architecture search. Second, be honest that on this data you cannot reliably detect a small accuracy improvement at all, which is itself an argument for the simpler, cheaper option. ## What would flip the decision - **Many related series.** This is the big one. Four thousand ATMs with 120 days each is half a million observations of the same underlying behaviour. A single global recurrent model pooled across them can learn payday effects, holiday effects and day-of-week shape once and share them, conditioning on per-ATM features for location and size. Cross-series learning, not sequence length, is what deep forecasters are actually for. The unit of data is the portfolio. - **Rich covariates and non-linear interactions.** If withdrawals depend jointly on paydays, holidays, weather and local events in ways that interact, a neural model with many input channels has something to offer that a small linear-in-structure model does not — but only once there is enough data to estimate those interactions. - **Long or multi-quantile horizons across many series.** One global model producing a full horizon and several quantiles for thousands of series at once can be cheaper to operate than thousands of individually fitted classical models. ## The organisational angle A neural forecaster for one series is a permanent operating cost — training infrastructure, a serving path, monitoring, someone who understands it — with no accuracy upside. The durable answer is a layered platform: seasonal-naive as the floor, a classical per-series model as the default that covers the short and sparse tail, and a pooled deep model introduced only where the portfolio is large enough to earn it, always benchmarked against the two below it on the same rolling origins. State the revisit trigger explicitly when you make the call — more series, new covariates, a longer horizon — so the decision is a dated engineering judgment rather than a veto on a technology.