Time-Series Forecasting Services
Forecasting GDP, inflation, demand, or prices means modelling a series’ own structure—its trend, seasonality, and dynamics—and being honest about uncertainty. MAS Research builds classical time-series forecasts, from ARIMA and exponential smoothing to dynamic factor and MIDAS models, evaluated out of sample and reported with proper prediction intervals.
Time-series forecasting predicts the future values of a series from its own past patterns—trend, seasonality, and autocorrelation. Classical methods include ARIMA and SARIMA, exponential smoothing (ETS), ARIMAX for external predictors, MIDAS for mixed-frequency data, and dynamic factor models for many related series. Good forecasting is judged by out-of-sample accuracy and reports uncertainty through prediction intervals, not just a point forecast.
What time-series forecasting does
Time-series forecasting predicts where a series is heading by modelling the structure in its own history—its trend, its seasonality, and its autocorrelation (how each value relates to recent ones). Unlike cross-sectional prediction, the ordering of the data is everything: the past is the main input, and the goal is an accurate, honestly-quantified view of the future. It answers the practical questions that drive economic and business planning—where are GDP, inflation, demand, or prices likely to go, and how confident can we be?
The classical toolkit is deep and well-understood. ARIMA models capture autocorrelation and trend through autoregressive and moving-average terms with differencing; SARIMA adds seasonality; ARIMAX brings in external predictors. Exponential smoothing (ETS) models—including Holt–Winters—weight recent observations more heavily and handle trend and seasonality elegantly, often forecasting surprisingly well. For richer settings, MIDAS models combine data sampled at different frequencies (using monthly indicators to forecast quarterly GDP, say), and dynamic factor models extract common factors from many related series to forecast efficiently in data-rich environments. For multivariate systems, VAR forecasting models several series jointly.
When to use which approach
Choose the method to match the series and the data. Use ARIMA/SARIMA for a single series with clear autocorrelation and (for SARIMA) seasonality; use ETS when trend and seasonality are the dominant features and a robust, well-calibrated forecast is wanted—it is a strong, hard-to-beat baseline. Use ARIMAX or a regression-with-ARIMA-errors approach when external drivers genuinely help. Use MIDAS when useful predictors arrive at a higher frequency than the target (classic for nowcasting). Use dynamic factor models when you have many correlated series and want to exploit their common structure, and VAR when several series forecast each other.
One principle governs all of them: a more complex model is not automatically a better forecaster. Simple methods frequently win, especially at longer horizons, and the only way to know is out-of-sample evaluation. We benchmark candidates against simple baselines and each other on data they have not seen, rather than selecting on in-sample fit—which rewards overfitting and misleads.
Choosing a time-series method
| Method | Best for | Note |
|---|---|---|
| ARIMA / SARIMA | Autocorrelated (and seasonal) single series | Workhorse; needs stationarity handling |
| Exponential smoothing (ETS) | Trend & seasonality; robust baselines | Often hard to beat |
| ARIMAX | When external predictors genuinely help | Needs future values of predictors |
| MIDAS | Mixed-frequency data & nowcasting | Uses high-frequency indicators |
| Dynamic factor models | Many related series (data-rich) | Exploits common factors |
What makes a forecast trustworthy
A credible forecast is built and judged by discipline, not by how good the fitted line looks on past data. The decisive step is out-of-sample evaluation: candidate models are compared on data held back from estimation, using time-series cross-validation (rolling or expanding windows) that respects the time ordering—never a random split, which would leak the future into the past. Accuracy is measured with appropriate metrics (RMSE, MAE, MAPE, and scaled measures like MASE that allow comparison across series), and—crucially—against simple benchmarks such as a naive or seasonal-naive forecast. A model that cannot beat the naive benchmark is not adding value, however sophisticated.
Two further disciplines matter. First, uncertainty must be reported: a point forecast alone is misleading, so we provide prediction intervals that widen with the horizon and reflect genuine forecast uncertainty, and check that their coverage is calibrated. Second, the usual time-series groundwork applies—stationarity and differencing for ARIMA, residual diagnostics to confirm the model has captured the structure (no leftover autocorrelation), and attention to structural breaks that can break a forecast. We report the evaluation protocol, the benchmark comparison, and the interval calibration, so the forecast can be trusted rather than just presented.
Judge forecasts out of sample, against simple benchmarks, with intervals. In-sample fit rewards overfitting; a model that cannot beat a naive or seasonal-naive forecast is not adding value. And a point forecast without a calibrated prediction interval hides the uncertainty that matters most for decisions.
Software
We deliver time-series forecasting in established, reproducible tools—R (forecast, fable, tsibble) and Python (statsmodels, pmdarima, sktime)—with automated and manual model selection, time-series cross-validation, benchmark comparison, calibrated prediction intervals, and residual diagnostics, all with versioned code.
How we deliver a forecasting study
Time-series forecasting sits within our wider forecasting practice—so the method fits the series, the evaluation is out of sample, and uncertainty is reported honestly.
We start from the series and the forecasting task—horizon, frequency, whether external drivers or many related series are available—and from the data’s features (trend, seasonality, breaks). We fit a set of candidate models from strong simple baselines (naive, ETS) up to ARIMA, ARIMAX, MIDAS, or factor models as the data justify, and compare them with time-series cross-validation against benchmarks.
Reporting sets out the data features, the candidate models, the out-of-sample evaluation and benchmark comparison, the chosen model and why, and the forecasts with calibrated prediction intervals—so the forecast and its uncertainty can both be judged.
You receive the forecasts with prediction intervals, the out-of-sample accuracy and benchmark comparison, the chosen model with diagnostics, scenario or external-driver variants where relevant, clear visualisations, and reproducible analytical code and analysis-ready files (where appropriate and permitted).
Series we forecast
Across macroeconomics, markets, and sectors—each series matched to a method suited to its structure.
Macroeconomic Indicators
GDP, inflation, unemployment, and other indicators—often with MIDAS nowcasting from higher-frequency data and factor models in data-rich settings.
Interest & Exchange Rates
Rates and currencies, where dynamics, external drivers, and honest uncertainty all matter.
Financial Markets
Stock-market and asset-return series—with realistic expectations about predictability and strong benchmarking.
Commodity & Energy
Commodity prices and energy demand, often seasonal and driven by external factors.
Financial Risk & Volatility
Volatility and risk forecasting, drawing on GARCH models for time-varying variance.
Demand & Business Series
Sales, demand, and operational series for planning—typically seasonal and well-served by ETS and ARIMA.
Time-series forecasting: common questions
Need a forecast you can defend?
Whether it is GDP, inflation, demand, or prices, we build the time-series model that fits your series, evaluate it out of sample against honest benchmarks, and report the forecast with calibrated prediction intervals—not a point estimate dressed up as certainty.