Time series forecasting is the practice of predicting future values of a variable indexed by time, using historical observations and statistical, machine learning, or deep learning models that capture trend, seasonality, and autocorrelation.
Time series forecasting is the discipline of predicting future values of a variable that is indexed by time, given a history of past observations. The defining feature is order, observations are not exchangeable, the value at time t depends on values at times t-1, t-2, and so on. This temporal structure is what separates time series forecasting from generic regression or classification, and it is why specialised models exist for it.
In a finance context the variable might be daily cash receipts for a single customer, weekly invoice volume by business unit, monthly DSO, or the EUR/USD rate. The goal is usually a forecast horizon of days to quarters, often with confidence intervals attached. Modern practice treats forecasting as a probabilistic problem, the output is a distribution over future values, not a single number.
Real-world series rarely look like clean signals. Treasury and FP&A teams need to recognise the components that show up in their data. Trend is the long-run direction, slow growth in receivables as the business scales, for example. Seasonality is regular repetition tied to the calendar, end-of-quarter payment spikes, summer slowdowns, or weekly Monday-heavy clearing patterns. Cyclicity is repetition that is not calendar-locked, such as the broader economic cycle that affects bad debt.
Beyond these, finance series often show level shifts (a major customer churns, the baseline drops), autocorrelation (yesterday's cash inflow tells you something about today's), and regime changes where the underlying process itself shifts, think COVID, rate-hiking cycles, or a switch from cheque to ACH. Any model that ignores these structures will either under-fit during stable periods or over-react during shocks.
There are four broad families in production use today. Statistical models include ARIMA, SARIMA, exponential smoothing (ETS), and state-space models. They are interpretable, cheap to train, and remain strong baselines for short horizons on well-behaved series. Machine learning approaches use gradient boosting libraries like XGBoost and LightGBM on engineered features, lags, rolling means, calendar dummies, and exogenous drivers, scaling well across thousands of related series.
Deep learning brings recurrent architectures (LSTM, GRU) and modern sequence models (Temporal Fusion Transformer, N-BEATS, DeepAR), which learn cross-series patterns and handle high cardinality natively. Foundation models such as TimesFM, Chronos, Moirai, and Lag-Llama are pretrained on billions of time points and offer zero-shot forecasting across new series.
Evaluating a forecaster is harder than evaluating a classifier because the data has temporal order. The gold standard is rolling-origin backtesting, train on data up to time t, forecast the next h steps, slide the origin forward, repeat. This mimics live deployment and exposes regime sensitivity.
For point forecasts, common metrics are MAE (mean absolute error), RMSE (root mean squared error), MAPE (mean absolute percentage error), and sMAPE for symmetry. For probabilistic forecasts, which output a distribution rather than a point, the right metrics are quantile loss (also called pinball loss) at chosen quantiles and CRPS (continuous ranked probability score). Finance teams that only track MAPE on a point forecast are flying blind on tail risk, which is precisely the risk treasury cares about.
Time series forecasting is the engine inside every cash flow forecasting product. Each customer's payment timing is a time series. Each invoice category's volume is a time series. Each FX pair, each disputed-invoice resolution lag, each chargeback rate, all time series. A 13-week cash flow forecast is a hierarchy of hundreds or thousands of underlying series that must be modelled, reconciled, and rolled up.
The choice of method has direct consequences. ARIMA on the top-line works for stable receipts but collapses when a major customer changes payment behaviour. Gradient boosting on per-customer features captures heterogeneity but needs thoughtful feature engineering. Deep learning models share strength across customers but demand GPU infrastructure. Foundation models offer zero-shot accuracy on long-tail customers where statistical models lack data, at the cost of inference latency. The right architecture for a rolling forecast depends on series count, horizon, accuracy target, and operating budget.
Transformance.ai applies modern time series methods including foundation models across AR-driven cash forecasting workflows, see our research notes on real-world accuracy benchmarks.
The most common failure mode is look-ahead bias, using information at training time that would not have been available in production. Close cousins are train-test leakage from shuffled splits and ignoring known-future variables like holiday calendars. Many teams also default to point forecasts when intervals are what treasury actually needs, the question is not what the cash position will be, it is how likely it falls below the covenant threshold.
Best practice today includes probabilistic forecasting with quantile outputs (P10, P50, P90), hierarchical reconciliation so customer-level forecasts sum coherently to business-unit and group level (top-down, bottom-up, or middle-out), explicit holiday calendars, and inclusion of exogenous variables such as billing volume, sales pipeline, or macro indicators. Combined with rigorous rolling backtests, these practices turn cash flow forecasting from a quarterly art project into a defensible, audit-ready process.
Regression treats observations as exchangeable and predicts a target from features. Time series forecasting respects temporal order, models autocorrelation, trend, and seasonality, and evaluates with rolling-origin backtests rather than random train-test splits.
There is no single winner. Statistical models like ARIMA and ETS are strong baselines for short horizons on stable series. Gradient boosting wins when you have many related series and good features. Deep learning and foundation models like TimesFM and Chronos lead on complex, high-cardinality problems but cost more to run. Benchmark on your own data.
Probabilistic forecasting outputs a full distribution over future values rather than a single point. Typical outputs are quantiles such as P10, P50, and P90. This matters for finance because covenant breaches, liquidity buffers, and risk decisions depend on the tails of the distribution, not just the median.
For point forecasts use MAE, RMSE, MAPE, or sMAPE depending on scale and symmetry needs. For probabilistic forecasts use quantile loss (pinball loss) at chosen quantiles and CRPS. Always evaluate using rolling-origin backtesting to mimic live deployment.
When forecasts exist at multiple levels (customer, segment, business unit, group) the sum of lower-level forecasts often does not match the top-level forecast. Hierarchical reconciliation methods (top-down, bottom-up, middle-out, MinT) adjust the forecasts so they are coherent across the hierarchy while preserving accuracy.
Cash flow forecasting is an application of time series forecasting to AR receipts, AP disbursements, FX flows, and other financial series. Each underlying series is forecast individually and then reconciled into a 13-week or rolling cash forecast. The choice of forecasting method directly drives accuracy and infrastructure cost.
