Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Mitigating Volatility Drag Using Machine Learning Models

Andersson, Douglas LU and Grahm, Carl Johan (2025) NEKH02 20251
Department of Economics
Abstract (Swedish)
We observe that, while predicting returns remain difficult and uncertain, volatility exhibit more persistence and predictability due to well-known patterns such as volatility clustering (Mandelbrot, 1967). We test whether short-horizon volatility forecasts can be used to time leverage in a passive index strategy so as to mitigate the growth penalty from volatility drag in leveraged exchange-traded funds (LETFs). Using OMXS30GI daily data (2012-2025), we train two ensemble models, Random Forest (RF) and Extreme Gradient Boosting (XGB), to forecast next-day volatility and map those forecasts to discrete leverage choices
{1x, 1.5x, 2x } via a log-growth break-even rule. We compare against static 1.5x and 2x benchmarks as well as an... (More)
We observe that, while predicting returns remain difficult and uncertain, volatility exhibit more persistence and predictability due to well-known patterns such as volatility clustering (Mandelbrot, 1967). We test whether short-horizon volatility forecasts can be used to time leverage in a passive index strategy so as to mitigate the growth penalty from volatility drag in leveraged exchange-traded funds (LETFs). Using OMXS30GI daily data (2012-2025), we train two ensemble models, Random Forest (RF) and Extreme Gradient Boosting (XGB), to forecast next-day volatility and map those forecasts to discrete leverage choices
{1x, 1.5x, 2x } via a log-growth break-even rule. We compare against static 1.5x and 2x benchmarks as well as an unlevered index proxy. Out of sample (2022-2025), the models allocate to 2x 77 - 85% of days, de-risking after drawdowns relative to static 2x exposure. The mechanism is consistent with theory: when signals arrive with latency, the quadratic
( k222 ) penalty dominates. We conclude that price-only volatility forecasts are insufficient for effective leverage timing on Swedish data; future work should incorporate anticipatory volatility measures, market-trend/direction and asymmetric switching rules. (Less)
Please use this url to cite or link to this publication:
author
Andersson, Douglas LU and Grahm, Carl Johan
supervisor
organization
course
NEKH02 20251
year
type
M2 - Bachelor Degree
subject
keywords
Machine learning, Random Forest, Volatility Drag, XGBoost, LETF, Leverage
language
English
id
9212648
date added to LUP
2025-12-08 08:36:02
date last changed
2025-12-08 08:36:02
@misc{9212648,
  abstract     = {{We observe that, while predicting returns remain difficult and uncertain, volatility exhibit more persistence and predictability due to well-known patterns such as volatility clustering (Mandelbrot, 1967). We test whether short-horizon volatility forecasts can be used to time leverage in a passive index strategy so as to mitigate the growth penalty from volatility drag in leveraged exchange-traded funds (LETFs). Using OMXS30GI daily data (2012-2025), we train two ensemble models, Random Forest (RF) and Extreme Gradient Boosting (XGB), to forecast next-day volatility and map those forecasts to discrete leverage choices 
{1x, 1.5x, 2x } via a log-growth break-even rule. We compare against static 1.5x and 2x benchmarks as well as an unlevered index proxy. Out of sample (2022-2025), the models allocate to 2x 77 - 85% of days, de-risking after drawdowns relative to static 2x exposure. The mechanism is consistent with theory: when signals arrive with latency, the quadratic 
( k222 ) penalty dominates. We conclude that price-only volatility forecasts are insufficient for effective leverage timing on Swedish data; future work should incorporate anticipatory volatility measures, market-trend/direction and asymmetric switching rules.}},
  author       = {{Andersson, Douglas and Grahm, Carl Johan}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Mitigating Volatility Drag Using Machine Learning Models}},
  year         = {{2025}},
}