Time series analysis and forecasting

No. Reading a trend or seasonality by eye leads to mistaking noise for signal and to projecting forward a pattern that may never have been there. This applies to any time-indexed data, such as demand, consumption, incidence or exposure, about which you want to claim more than a visual impression. A time series has trend, seasonality and dependence, and each component is assessed with appropriate methods: whether a series rises, falls or seems to repeat every year is not read off the curve, it is tested. Inspecting the chart does not separate multiple or overlapping seasonalities, nor does it distinguish a real trend from a short-term fluctuation. Characterising the structure properly is what turns a visual reading into a defensible claim and, on that basis, into a forecast with an interval rather than an extrapolation by hand. What the service provides is the diagnostic evidence behind each component interpreted, not an interpretation of the shape of the curve.

By separating the structure of the series into its components with formal methods, not by visual inspection. This applies to any series whose trend, seasonality or cycles you want to characterise before modelling or drawing conclusions. The underlying trend is estimated using regression, smoothing and filters, separating it from short-term fluctuation; stationarity is tested with unit root tests such as augmented Dickey-Fuller and KPSS; and the signal is decomposed through Fourier analysis, the periodogram and STL decomposition to identify its component frequencies. That spectral decomposition reveals multiple or overlapping seasonalities that the chart cannot separate. When the series is not homogeneous across its whole span, changes in level or slope at breakpoints are also tested. What you receive is a characterisation of each component with the relevant diagnostics and tests, so that saying "there is a trend" or "there is annual seasonality" rests on a test, not on a reading of the curve.

ARIMA is one of the models, not the only one: the model is chosen by the structure of the data, and ARIMA is not applied to everything by default. This applies to the forecasting stage, once the series has been characterised. A series with a single regular seasonality, one with several overlapping seasonalities and one with regime shifts call for different models. That is why dependence, seasonality and stationarity are assessed first, and the choice is made between ARIMA and SARIMA for temporal dependence and regular seasonality, exponential smoothing and state space models (ETS), TBATS for multiple seasonalities, and Prophet or other flexible methods when the structure of the series and the forecasting goal justify them. The choice is justified by what the data support, not by a preference for a tool. Applying the same model to any series is one of the most common mistakes in forecasting. What you receive is the model that the structure of the series justifies, not a single template for every series.

A horizon with its prediction intervals, not a single number: a forecast without a margin of error is a figure with no defence. This applies to any projection that will be used for planning, because deciding on a deceptively precise point means assuming a certainty that the data do not provide. The forecast always comes with intervals that quantify uncertainty and widen with distance in time: the further ahead, the wider the interval, because uncertainty accumulates. That lets you plan on a range consistent with the model and the observed history rather than on a single figure. An honest limit applies: the interval reflects the estimated uncertainty under the model and the observed structure; it cannot ensure that any future regime shift, or an event without precedent in the series, will be captured. Reliability also rests on out-of-sample validation, not on how well the model fits the past. What you receive is a horizon with its uncertainty made explicit, not a promise to hit the exact value.

By validating it against the series' own history while respecting the order of time, not by how well it reproduces the data already known. This applies to any forecasting model, because a model that fits the past perfectly can forecast the future very badly: fitting and forecasting are not the same thing. Validation is done on rolling windows that respect the direction of time, without contaminating the fit with future information, and models are compared by their out-of-sample forecast error, not by their goodness of fit to the data they had already seen. Residual diagnostics are added to this, checking that the model has captured the structure and leaves no signal unexplained. The measure of quality is the forecast, not the fit. That honest temporal validation is what helps distinguish a model whose predictive ability has been tested retrospectively from one that merely reproduces the past well and will fail as soon as the future departs from it.

Not as the main purpose of the service: this service characterises the structure of a series and projects it; it does not monitor the series to raise alerts when a value deviates from what is expected. This limit is worth stating because "analysing a series" can be understood either way. Operational detection of anomalies or aberrations, which means comparing each new value against an expected value to trigger an alert, is the job of surveillance, an approach geared towards early warning rather than characterisation and forecasting. Here, by contrast, the focus is on understanding the structure of the series (trend, seasonality, dependence) and projecting its evolution with intervals. A good model of the series can be the foundation on which a detection system is later built, but that alert system is a separate piece of work. If the need is to monitor the series to detect deviations, that is surveillance; if it is to characterise and forecast it, that is this service.

Not directly: attributing a change in the series to the effect of an intervention is a causal question, different from characterising and projecting the series, which is what this service does. This applies when what you are after is not a forecast of how the series will evolve, but an estimate of whether a change point (a policy, a protocol, an event) altered its course. That analysis, using interrupted time series, synthetic controls and other quasi-experimental designs, belongs to causal inference or to impact evaluation, depending on the design and the context, because it requires building a counterfactual: what would have happened without the intervention. This service can test descriptively whether there is a change in level or slope at a breakpoint, but describing that change is not the same as attributing it causally to an intervention. If the question is "what effect did this intervention have on the series?", the framing is causal inference; if it is "how does the series behave and where is it heading?", it is this service.

You need the time-indexed series, meaning the indicator recorded at intervals with its timestamp, and the sampling should be as regular as possible, although handling gaps and missing values without breaking the temporal structure is part of the preliminary work. This applies to series from any domain, such as demand, consumption, incidence or exposure, because the methodological core of characterising the signal and forecasting is the same whatever the field. The client records and measures the series; the service checks the regularity of the sampling, characterises the structure, models, forecasts and validates. What shapes the scope is not the domain but the quality and length of the series: a short series or one with many gaps limits what can be estimated and projected, and we say so honestly before modelling. What you receive is the characterisation of the series and, where appropriate, its forecast with intervals, based on data that the client records and keeps.