Neural Network Methods for Gas Storage Valuation: A Least Squares Monte Carlo Study with Application to Gas Storage Denmark
Translated title
Neural Network Methods for Gas Storage Valuation
Author
Sønder, Rasmus Flaathen
Term
4. term
Education
Publication year
2026
Submitted on
2026-05-27
Abstract
This thesis studies how to value natural gas storage contracts by treating the decision to inject or withdraw gas as a sequential decision-making problem under uncertainty. We use Least Squares Monte Carlo (LSMC), a simulation-based method that estimates the continuation value (the expected future value from waiting) with six approaches: polynomial OLS regression, spline regression, and nodewise and joint-input variants of two neural networks, a multi-layer perceptron (MLP) and a Kolmogorov–Arnold Network (KAN). Spot price dynamics are modeled with a one-factor mean-reverting model calibrated to Trading Hub Europe prices, and the framework is applied to the Danish storage system operated by Gas Storage Denmark (GSD) and to a stylized Dutch salt cavern. In the tested setups, polynomial OLS and splines outperform both neural network approaches. For GSD, OLS yields €123.481M, while the best neural network (the joint-input MLP) reaches €107.101M, a gap of about 13%. The continuation value depends almost linearly (near-affine) on inventory, which favors polynomial regression and limits the benefits of larger networks. The study is limited by the one-factor price model, the cost of training neural networks, and the computational inefficiency of the nodewise KAN, which is roughly 600 times slower than OLS. Overall, the results indicate that classical LSMC with polynomial regression is both more accurate and orders of magnitude faster than neural network alternatives for the class of gas storage problems studied.
Denne afhandling undersøger, hvordan naturgas-lagerkontrakter kan værdiansættes ved at se beslutningen om at fylde op eller tømme som et sekventielt beslutningsproblem under usikkerhed. Vi bruger Least Squares Monte Carlo (LSMC), en simuleringsbaseret metode, der estimerer fortsættelsesværdien (den forventede fremtidige værdi ved at vente) med seks tilgange: polynomiel OLS-regression, spline-regression samt nodevise og fælles-input varianter af to neurale netværk, multi-layer perceptron (MLP) og Kolmogorov–Arnold Network (KAN). Spotpriser modelleres med en en-faktor middelreversionsmodel kalibreret til Trading Hub Europe, og rammeværket anvendes på det danske lagersystem hos Gas Storage Denmark (GSD) og en stiliseret hollandsk saltkaverne. I de undersøgte opsætninger klarer polynomiel OLS og splines sig bedre end de neurale netværk. For GSD giver OLS €123.481M, mens det bedste neurale netværk (MLP med fælles input) når €107.101M, en forskel på cirka 13 %. Den næsten lineære (nær-affine) afhængighed mellem fortsættelsesværdi og lagerbeholdning favoriserer polynomiel regression og begrænser gevinsten ved større netværk. Studiet er begrænset af den en-faktor prismodel, omkostningerne ved at træne neurale netværk og den lave beregningsmæssige effektivitet for nodevis KAN, som er omkring 600 gange langsommere end OLS. Samlet peger resultaterne på, at klassisk LSMC med polynomiel regression både er mere præcist og væsentligt hurtigere end neurale alternativer for de typer gaslagerproblemer, der er undersøgt.
[This abstract has been rewritten with the help of AI based on the project's original abstract]