PhD Researchers & Doctoral Candidates
A properly evaluated forecasting study for a thesis—benchmarked out-of-sample, not just fit in-sample.
Research Methods
Forecasting & Predictive Analytics
Any model can fit the past. What matters is whether it predicts the future—and the only honest way to know is to test it on data it has never seen. We build forecasts for economic and financial series that are validated out-of-sample, benchmarked against real alternatives, and reported with their uncertainty.
ARIMA · VAR · MIDAS · factor models
Bayesian & ML forecasting
Evaluated out-of-sample
Reproducible, journal-ready
Sample output
history
forecast + interval
Overview
Forecasts judged by what they get right
A model that fits history perfectly can still forecast badly—overfitting rewards in-sample accuracy and punishes out-of-sample performance. That is the trap most forecasting papers fall into: an impressive R-squared on data the model was estimated on, and no evidence it works on anything else. The discipline that matters is honest out-of-sample evaluation, and it is where we start.
We match the method to the series and the horizon, then let performance decide. A single seasonal series may call for SARIMA or exponential smoothing; several related indicators for a VAR or dynamic factor model; a target driven by higher-frequency data for MIDAS. Where non-linearities matter, machine- and deep-learning methods, or a hybrid that combines an econometric backbone with an ML component—but only when they beat the simpler benchmark on a fair test.
Every forecast is evaluated on held-out data against benchmarks including a naive one, reported with prediction intervals rather than as false-precision point estimates, and delivered with reproducible code—so the forecast is defensible whether it goes into a paper or a decision.
Who We Work With
For anyone who needs to predict a series
If a decision or a paper turns on where a number is headed, this is the right desk to write to.
Economics & Finance Researchers
Macro and market forecasting to the standard the field expects, with honest predictive-accuracy testing.
Central-Bank & Policy Teams
GDP, inflation, and indicator forecasting where prediction intervals and model comparison inform real decisions.
Energy & Commodity Analysts
Demand and price forecasting for energy and commodity markets, including mixed-frequency drivers.
Research Institutes & Think Tanks
Applied forecasting for economic-outlook and scenario work that has to be defensible.
Corporates & Industry R&D
Demand, risk, and indicator forecasting translated from academic rigor into planning decisions.
Capabilities
The full forecasting toolkit & what we forecast
Organized by method family, with the targets we work on most. If your series needs an approach not listed here, ask—this is the core, not the boundary.
Univariate & multivariate
The workhorse forecasting models, from seasonal univariate methods to multivariate and mixed-frequency approaches.
- ARIMA
- SARIMA
- ARIMAX
- VAR forecasting
- Exponential smoothing
- MIDAS
- Dynamic factor models
Uncertainty & non-linearity
Methods for full predictive densities, volatility forecasting, and the non-linear patterns classical models miss.
- Bayesian forecasting
- GARCH forecasting
- Machine-learning forecasting
- Deep-learning forecasting
- Hybrid econometric-ML models
Macro, market & sector series
The economic and financial targets we most often model—each with method choices suited to it.
- GDP
- Inflation
- Interest rates
- Exchange rates
- Stock markets
- Commodity prices
- Energy demand
- Unemployment
- Financial risk
- Economic indicators
How the Analysis Works
Six steps from series to validated forecast
A transparent sequence organized around one principle: the forecast is judged on data the model never saw. Nothing is a black box.
Steps are adapted to your target: frequency, horizon, history length, and whether interpretability or raw accuracy is the goal. We confirm the approach with you before building.
-
1
Frame
Define the target, the horizon, and the loss that matters—a one-step point forecast and a twelve-step density are different problems.
Inputs: target · horizon · frequency · loss
-
2
Explore
Examine trend, seasonality, breaks, and stationarity, and prepare the series and any predictors.
Checks: trend · seasonality · breaks · stationarity
-
3
Model
Fit a set of candidate models—including a naive benchmark—that the data's features justify.
Models: ARIMA · VAR · factor · Bayesian · ML · hybrid
-
4
Evaluate
Compare forecasts out-of-sample with rolling or expanding windows and appropriate loss functions.
Metrics: RMSE · MAE · rolling windows · Diebold-Mariano
-
5
Quantify
Produce prediction intervals or predictive densities so the uncertainty around the forecast is explicit.
Output: intervals · densities · fan charts
-
6
Report
Deliver forecasts, evaluation tables, figures, methodology, and reproducible code you keep.
Output: forecasts · evaluation · methods · R/Python code
Rigor by default
The checks that separate a forecast from a fit
In-sample accuracy proves nothing about the future. The out-of-sample discipline that makes a forecast credible is standard on every engagement.
Included on every project
- Out-of-sample evaluation on held-out data
- Comparison against a naive and rival benchmarks
- Tests of predictive accuracy (Diebold-Mariano)
- Prediction intervals, not point forecasts alone
- Reproducible, versioned code you keep
What You Receive
Every engagement, delivered in full
Not a black-box result and a number, but a complete, documented package you can submit, defend, and reproduce.
- Clean, documented series and predictor dataset
- Candidate models with selection rationale
- Out-of-sample evaluation against benchmarks
- Point forecasts with prediction intervals
- Forecast plots and fan charts
- Interpretation of the forecast and its drivers
- Reproducible R (fable) or Python code
- Journal-ready methodology and results sections
- Technical responses to methodological reviewer comments, where required
Where this fits
Part of a larger arc
Forecasting is strongest when the time-series foundations ahead of it are sound and the reporting after it is precise—each handled with the same care.
Research Design & Planning
Identification strategy, power, and specification decided before estimation begins.
Explore methodsStatistical & Methodological Audit
An independent check of assumptions, specification, and reproducibility before submission.
Explore auditPublication & Research Support
Methods and results reporting, journal selection, and reviewer-response support.
Explore support
FAQ
Common questions
Answers to what most researchers and project leads ask before we begin a forecasting engagement.
By the series and the horizon, tested rather than assumed. A single series with clear seasonality may be well served by SARIMA or exponential smoothing; several related indicators by a VAR or dynamic factor model; a target driven by higher-frequency data by MIDAS. We benchmark candidate models out-of-sample and let predictive performance, not preference, decide.
Not always—and that is exactly why we test it. ML and deep-learning methods can capture non-linearities that ARIMA cannot, but on many economic series a well-specified classical or hybrid model matches or beats them, especially with limited data. We compare them honestly on the same out-of-sample task rather than assuming the newer method wins.
Out-of-sample, always. We hold out data the model never sees, compare forecasts against benchmarks (including a naive one) using loss functions such as RMSE and MAE, and apply tests of predictive accuracy like Diebold-Mariano. In-sample fit alone is not evidence a forecast will work.
Yes. A point forecast without a sense of its uncertainty is of limited use for a decision. We produce prediction intervals or full predictive densities—particularly with Bayesian methods—so the range of plausible outcomes is explicit.
Yes. MIDAS and related approaches let a lower-frequency target (say, quarterly GDP) be forecast using higher-frequency predictors (monthly or daily indicators) without discarding information by aggregating everything to the slowest frequency.
Yes, and it is often the strongest approach. Combining a structural or econometric model with a machine-learning component—for example, modeling the linear structure classically and the residual non-linearity with ML—can outperform either alone, and we validate the combination out-of-sample.
R (forecast, fable, and related packages) and Python (statsmodels, and deep-learning frameworks where needed), with EViews where preferred. You receive versioned code, the evaluation results, and a methods section written to journal standards.
Yes, and it is the best time. The data frequency, the horizon, the amount of history, and whether interpretability or pure accuracy matters all shape the right method—settling them up front prevents building a project around a model the data cannot support.
Related Expertise
Often paired with econometrics
Need to predict a series?
Tell us the target, the horizon, and the data you have—we'll tell you the right method, and prove it out-of-sample before you rely on it.