Skip to content
CONKLING KOPCIKECONOMICS

CK Forecasting

How the forecasts are made and scored

Every number on the forecast pages comes from the steps below, run each day after the IMF data update. The code and the archive of past IMF editions are public in the site’s repository.

What is forecast

Real GDP growth, inflation (annual average consumer prices) and the unemployment rate, on the IMF World Economic Outlook’s definitions, for the economies with a full country profile. The horizons are the three years after the last year the latest IMF edition reports: from an October edition, that is the edition’s own year (mostly past by then, so close to a nowcast), the next year and the one after.

These are baseline forecasts: CK’s central expectation from what is known now. They are not scenarios. The simulator answers “what if” questions with illustrative paths, which are never scored or shown as forecasts (see the research standards).

Data and vintages

Each forecast starts from one IMF World Economic Outlook edition: its reported history and its own forecast. The backtest uses every earlier edition of the same season (April or October) since April 2008, each exactly as it was published at the time, from DBnomics’ archive of the IMF’s releases and from the editions CK has saved since. No origin sees data published after it: not revised history, not later outcomes.

Outcomes are the latest edition’s reported figures. The European Union is scored only from April 2020, when the IMF’s figures stopped including the United Kingdom.

The models

  • ARIMA(1,0,0): the series reverts towards its average over the last 25 years, at the speed its own history shows.
  • ETS(A,N,N), simple exponential smoothing: a level that follows recent years, projected flat.
  • The IMF’s forecast from the same edition, used as published.

Inflation is modelled as the logarithm of one plus the rate, so a hyperinflation year doesn’t swamp the fit. Both statistical models are fitted to each edition’s history as it was then.

Combining them

The forecast is a weighted average of the three. For each indicator and horizon, the weights (each between 0 and 1, adding up to 1, in steps of 0.05) are those that would have minimised the average error, across every country, of the forecasts made from earlier editions. Errors are divided by each series’ typical yearly change so that countries can be pooled. Weights use only forecasts whose outcomes the edition being forecast from already reported, so the backtest’s weights are what CK could have known at the time.

In practice the IMF’s forecast carries most of the weight, and the statistical models adjust it at longer horizons. Each forecast page shows its weights.

The 80% range

The range is the 10th to 90th percentile of the errors the same weighted average made over the same earlier forecasts, scaled to the series and placed around the forecast. It is checked out of sample: across countries, the share of later outcomes inside it must be between 65% and 92% at that horizon, or the horizon is withheld everywhere.

The scorecard

Each series’ scorecard shows, at each horizon, the mean absolute error, in percentage points, of CK’s forecast, of naive (no change from the last reported year) and of the IMF’s forecast, over every edition that can be scored, and the share of outcomes inside the 80% range. The first editions only train the weights and are not scored. Past forecasts are listed next to their outcomes: rows marked “Backtest” are reconstructed from each edition as published; rows marked “Published at the time” are what CK published then, kept unchanged.

When a forecast is published

  • At least 10 past editions scored at that horizon;
  • CK’s mean absolute error at least 5% below naive’s;
  • the range passes its coverage check.

Horizons are published from the first: a horizon that fails withholds it and every later one. A series whose first horizon fails is withheld, and its scorecard stays public with the reason. Beating the IMF is not required, and the scorecard shows where CK’s error is larger.

Limitations

  • About 15 editions per series is a short record: one error in a crisis year (2009, 2020) moves a scorecard, and the range is wide for that reason.
  • Annual data only: monthly and quarterly figures released since the IMF’s edition are not used yet.
  • The IMF’s forecast is the main input, so CK’s forecasts share its errors; they are a transparent, scored forecast, not an independent view.
  • Every method change gets a new version name (shown on each forecast) and a new backtest of the whole history.

← The forecasts