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
Pages
100
Abstract
This thesis studies how to value natural gas storage contracts by framing it as decision‑making over time under uncertainty (a stochastic dynamic optimization problem). It uses a simulation‑based approach, Least Squares Monte Carlo (LSMC), and compares six ways to approximate the continuation value—the expected value of waiting to act rather than trading now: polynomial OLS regression, spline regression, and nodewise and joint‑input variants of two neural network types, the multi‑layer perceptron (MLP) and the Kolmogorov–Arnold Network (KAN). Spot price dynamics are modeled with a one‑factor mean‑reverting process (tending toward a long‑run average), calibrated to Trading Hub Europe prices. The framework is applied to the Danish storage system of Gas Storage Denmark (GSD) and a stylized Dutch salt cavern. Results show that polynomial OLS and spline regression outperform both neural network approaches in the configurations studied. In the GSD case, OLS yields €123.481M, while the best neural network, the joint‑input MLP, reaches €107.101M—about a 13% gap. The near‑affine (almost linear) dependence of the continuation value on inventory favors polynomial regression and limits the benefits of larger networks. The study is limited by the one‑factor price model, the training cost of neural networks, and the computational inefficiency of nodewise KAN, which is roughly 600 times slower than OLS. Overall, the findings 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 man værdiansætter naturgaslagringskontrakter, ved at formulere problemet som beslutningstagning over tid under usikkerhed (stokastisk dynamisk optimering). Den anvender et simulationsbaseret rammeværk, Least Squares Monte Carlo (LSMC), og sammenligner seks måder at tilnærme den såkaldte fortsættelsesværdi – den forventede værdi af at fortsætte med at lagre gas frem for at handle nu: polynomiel OLS‑regression, spline‑regression samt nodevise og fælles‑input varianter af to neurale netværkstyper, multi‑layer perceptron (MLP) og Kolmogorov–Arnold‑netværk (KAN). Spotprisernes udvikling modelleres med en én‑faktor middeltilbagevendende proces (der tenderer mod et langsigtet gennemsnit), kalibreret til Trading Hub Europe‑priser. Rammeværket anvendes på det danske lagersystem hos Gas Storage Denmark (GSD) og en stiliseret hollandsk saltkaverne. Resultaterne viser, at polynomiel OLS og spline‑regression klarer sig bedre end begge neurale netværk i de undersøgte opsætninger. I GSD‑casen giver OLS €123.481M, mens det bedste neurale netværk, MLP med fælles input, når €107.101M – et gab på ca. 13 %. Den næsten‑affine (næsten lineære) afhængighed mellem fortsættelsesværdi og lagerbeholdning begunstiger polynomiel regression og begrænser udbyttet af større netværk. Studiet er begrænset af én‑faktor pris‑modellen, træningsomkostningerne for neurale netværk og den beregningsmæssige ineffektivitet ved nodevis KAN, som er omkring 600 gange langsommere end OLS. Samlet peger resultaterne på, at klassisk LSMC med polynomiel regression er både mere præcist og størrelsesordener hurtigere end neurale alternativer for de undersøgte gaslagerproblemer.
[This apstract has been rewritten with the help of AI based on the project's original abstract]