This project estimates UK quarterly GDP growth before the official figure is released, using monthly indicators (Index of Services, Index of Production) that come out well ahead of the quarterly number. I tried two approaches: a bridge equation and a Dynamic Factor Model (DFM) that handles monthly and quarterly data together. I also caught a misleading result along the way, which is worth reading before the headline numbers below.
| Model | Out-of-sample R² |
|---|---|
| Bridge equation | 0.425 |
| Dynamic Factor Model (rolling) | 0.488 |
The DFM beats the bridge equation, but only modestly, once both are tested properly on data they haven't seen.
My first pass at evaluating the DFM gave an R² of 0.951. That number was misleading — it came from checking the model against data it had already seen during training, using a single long-range forecast that flattened out almost immediately instead of updating as new data came in. Once I refit the model using training data only and tested it one quarter at a time on data it hadn't seen, the honest R² dropped to 0.488. Full detail on both versions is in the methodology doc.
DFM vs bridge equation, out-of-sample:
Neither model captures the full size of the 2020 GDP swing, but the DFM reacts more than the bridge equation at the peak, which is where its small edge comes from. From 2022 onward, in calmer conditions, the two models track actual GDP almost identically.
Full methodology and evaluation details are in docs/methodology.md.
- Quarterly target: UK GDP quarter-on-quarter growth (ONS series IHYQ)
- Monthly indicators: Index of Services and Index of Production, from ONS's GDP monthly estimate dataset
- Sample: January 1997 onward, the earliest point monthly GDP data exists
- All series converted to growth rates before modeling, since the levels are non-stationary
- Clone this repo
- Create a virtual environment and run
pip install -r requirements.txt - Open
notebooks/01_nowcasting_exploration.ipynband run all cells
data/raw - original ONS CSVs (monthly indicators, quarterly GDP) notebooks/ - main analysis notebook results/figures/ - saved plots docs/ - full methodology write-up
- Several of the 24 rolling DFM refits didn't fully converge within the maximum EM iterations, which adds some uncertainty to individual quarterly estimates
- The DFM doesn't report standard errors here, so I can't assess statistical significance the way I could for the bridge equation
- Only about 117 quarterly observations total, which limits how much weight any single evaluation split can carry
- Neither model captures the full size of extreme shocks like COVID, only a partial reaction to them