History in, future out — and the evaluation design is the hard part.
Forecasting looks like regression but breaks most of regression's assumptions. Observations are ordered and correlated, the test set must come after the training set, and the thing you are predicting is often several steps ahead rather than one.
The field has recently acquired a genuinely new option. Pretrained foundation models for time series produce useful forecasts on series they have never seen, with no training at all — the same zero-shot idea that arrived in NLP a few years earlier. They do not always beat a well-tuned classical model, but they remove the cold-start problem entirely.
Sub-categories within this task. The distinction that matters is not which is best, but which fits your constraints.
ARIMA / SARIMAETSThetaProphet
LightGBM / XGBoost with lag and rolling features
DeepARN-BEATSTemporal Fusion TransformerPatchTST
Chronos / Chronos-BoltTimesFMMoirai
The third line in each is the part that makes it harder than the textbook version.
Project net cash positions across currencies and entities to manage buffers.
The wrinkleRegulatory liquidity requirements make the tails matter more than the central forecast. You need calibrated quantiles, not a point estimate — and the model must be defensible to treasury and to supervisors.
Forecast withdrawals per location to schedule replenishment.
The wrinkleThousands of related series with strong local calendar effects — paydays, holidays, local events. A classic case for a single global model across all locations rather than one model per ATM.
Project balances under rate scenarios for asset-liability management.
The wrinkleThis is scenario forecasting, not point forecasting. The model must respond sensibly to rate paths it has never seen, which is exactly where purely learned models are weakest and structural models earn their place.
Forecast call and chat volume by interval to schedule staff.
The wrinkleIntraday seasonality on top of weekly and annual patterns, plus shocks from outages and product events. Hierarchical reconciliation across intervals, days and sites matters as much as the base forecast.
Named functions and the arguments that actually change the outcome. Comments flag the traps.
An expanding-window backtest is the only honest evaluation. A single train/test split tells you almost nothing.
import numpy as np
import pandas as pd
def backtest(series, model_fn, horizon=28, folds=6, step=28):
"""Expanding window: train on everything up to a cutoff, forecast
the next `horizon`, walk forward, repeat."""
results = []
for k in range(folds):
cut = len(series) - (folds - k) * step
train, actual = series[:cut], series[cut:cut + horizon]
if len(actual) < horizon:
continue
pred = model_fn(train, horizon)
# MASE: scaled against a seasonal naive baseline, so a value
# above 1 means you are worse than "same as last week".
naive = np.mean(np.abs(np.diff(train, 7)))
results.append({
"fold": k,
"MAE": np.mean(np.abs(actual - pred)),
"MASE": np.mean(np.abs(actual - pred)) / naive,
})
return pd.DataFrame(results)
# ALWAYS produce the naive baseline first. Plenty of deep learning
# forecasts lose to "same value as last week" - better to know early.
def seasonal_naive(train, h, period=7):
return np.array([train[-period + (i % period)] for i in range(h)])No training. Point the model at a series and it forecasts — with quantiles, which is what you need for liquidity work.
import torch
import numpy as np
from chronos import ChronosPipeline
pipe = ChronosPipeline.from_pretrained(
"amazon/chronos-bolt-base",
device_map="cuda", torch_dtype=torch.bfloat16,
)
context = torch.tensor(history_values, dtype=torch.float32)
# Returns SAMPLES, not a point forecast. The distribution is the point.
forecast = pipe.predict(context, prediction_length=28, num_samples=100)
q = np.quantile(forecast[0].numpy(), [0.05, 0.5, 0.95], axis=0)
low, median, high = q
# For treasury work the 5th percentile is often the number that matters,
# not the median - plan against the bad case.
print("median day 1:", median[0], " 5th pct:", low[0])Fits classical, boosted and deep models plus foundation models, then ensembles by backtest performance. A strong baseline in one call.
from autogluon.timeseries import TimeSeriesDataFrame, TimeSeriesPredictor
# Long format: one row per (series, timestamp)
data = TimeSeriesDataFrame.from_data_frame(
df, id_column="atm_id", timestamp_column="date"
)
predictor = TimeSeriesPredictor(
prediction_length=28,
target="withdrawals",
eval_metric="MASE",
known_covariates_names=["is_holiday", "is_payday"], # known in advance
quantile_levels=[0.05, 0.1, 0.5, 0.9, 0.95],
).fit(data, presets="high_quality", time_limit=1800)
print(predictor.leaderboard(data)) # see what actually won
preds = predictor.predict(data, known_covariates=future_calendar)Any cross-validation that mixes time order leaks the future. Use expanding or rolling windows exclusively. This is the single most common error in forecasting projects.
Seasonal naive — 'the same as this day last week' — is surprisingly strong. Report MASE so the comparison is explicit. A sophisticated model that loses to it is a finding, not a failure to hide.
Holidays and paydays are known in advance; competitor pricing is not. Using a covariate you will not have at prediction time is leakage, and it is easy to do accidentally.
For anything feeding a buffer or a limit, the quantiles are the deliverable. A point forecast with no interval cannot support a decision about how much cash to hold.
Branch forecasts should sum to regional, and regional to national. Independent models will not, and stakeholders notice immediately. Reconciliation is a separate step you must plan for.