Summary
This paper compares the ESMA methodology with a LASSO BIC Machine Learning model for predicting one-year future returns of European passive stock ETFs. The ESMA model relies solely on past return data, while the LASSO BIC model incorporates macroeconomic variables such as GDP, implied volatility, interest rates and inflation.
To evaluate forecast performance, three dimensions are used.
- A forecast is considered unbiased when it does not systematically overestimate or underestimate returns, meaning that forecast errors are on average close to zero.
- Accuracy reflects how close forecasts are to realised returns, where smaller errors indicate better performance.
- Efficiency relates to whether all relevant available information is incorporated in the forecast. If not, there is still room to improve predictions.
The results show that the ESMA model produces forecasts that are unbiased, but inaccurate and inefficient. The LASSO BIC model produces more accurate forecasts, but these forecasts are biased and still inefficient. Comparative tests show that the LASSO BIC model is significantly more accurate and less inefficient than the ESMA model, despite producing higher bias.
The paper also finds that ETF characteristics matter. Larger ETFs initially produce more accurate forecasts, but this effect turns negative beyond a certain threshold. ETFs with higher liquidity risk produce less accurate but more unbiased forecasts, while ETFs with higher market risk generate more biased forecasts.
The paper concludes that a one size fits all approach to performance scenario forecasting is limited. Investors should carefully interpret ESMA based forecasts, while regulators may consider forecasting models that incorporate macroeconomic variables and account for portfolio specific characteristics.
