Forecasts
Experimental forecast of next year's financial stress index, based on both indices in the current year. All models are shown with their actual past errors.
Updated : 6 November 2026
Key results
1 2026 forecasts
The dot is the forecast; the bar spans the 10th to 90th percentiles of the model's past errors. The index uses a 0–100 scale (100 = highest level over 1998–2025).
2 Model performance
| # | Model | RMSE | MAE | Diebold-Mariano (p) |
|---|---|---|---|---|
| 1 | Naive model (previous year's value) | 13.2 | 10.7 | — |
| 2 | Random forest (machine learning) | 16.2 | 13.1 | 0.201 |
| 3 | Gradient boosting (machine learning) | 19.4 | 16.2 | 0.134 |
| 4 | Support vector regression (SVR) | 19.8 | 16.5 | 0.121 |
| 5 | Linear regression | 24.5 | 20.5 | 0.020 |
Out-of-sample evaluation over 17 years: each year is forecast using only earlier data. The Diebold-Mariano test compares each model with the naive one; a p-value below 0.05 signals a significant accuracy gap.
3 Forecast history and observed values
Year-by-year table
| Year | Observed | Naive | Linear | SVR | Forest | Boosting |
|---|---|---|---|---|---|---|
| 2026 | — | 50.0 | 99.4 | 67.0 | 57.6 | 53.1 |
| 2025 | 50.0 | 50.5 | 83.7 | 67.9 | 63.7 | 66.1 |
| 2024 | 50.5 | 76.1 | 82.2 | 70.3 | 75.8 | 71.9 |
| 2023 | 76.1 | 82.5 | 90.7 | 69.7 | 75.7 | 76.5 |
| 2022 | 82.5 | 62.4 | 67.2 | 68.7 | 62.0 | 60.9 |
| 2021 | 62.4 | 82.2 | 82.8 | 70.8 | 80.1 | 82.2 |
| 2020 | 82.2 | 87.8 | 84.0 | 70.6 | 85.5 | 90.1 |
| 2019 | 87.8 | 65.7 | 59.0 | 66.8 | 60.7 | 59.6 |
| 2018 | 65.7 | 56.3 | 97.7 | 68.8 | 61.2 | 64.6 |
| 2017 | 56.3 | 59.3 | 74.3 | 69.0 | 64.1 | 67.1 |
| 2016 | 59.3 | 64.6 | 59.1 | 68.8 | 56.6 | 42.9 |
| 2015 | 64.6 | 69.6 | 68.3 | 70.0 | 70.0 | 73.1 |
| 2014 | 69.6 | 84.0 | 56.1 | 71.5 | 60.3 | 54.0 |
| 2013 | 84.0 | 100.0 | 71.4 | 70.2 | 84.9 | 99.6 |
| 2012 | 100.0 | 89.3 | 55.2 | 69.4 | 75.1 | 75.4 |
| 2011 | 89.3 | 75.2 | 84.1 | 57.6 | 71.3 | 71.6 |
| 2010 | 75.2 | 78.7 | 45.8 | 42.9 | 64.6 | 78.4 |
| 2009 | 78.7 | 77.8 | 36.3 | 38.4 | 48.5 | 32.8 |
Methodology and data
Target: financial stress index in year t. Predictors: financial stress and climate risk indices in year t−1. Hyperparameters are fixed in advance and never tuned after seeing results: RBF-kernel SVR (C = 1, ε = 0.1); random forest of 500 depth-3 trees; gradient boosting of 200 depth-2 trees, learning rate 0.05. Tree models do not extend a trend: their forecast stays within the range of values already observed. Everything is recomputed automatically when a new year is added.
This is a statistical exercise on about thirty annual observations. It is not an official forecast, and no result establishes a causal relationship.