Risks of AI-Assisted Forecasted Inputs in Discounted Cash Flow Valuation: Reproducibility and Transparency

Romana ČIŽINSKÁ – Pavel NESET

https://doi.org/10.18267/pr.2026.vol.2587.7

Abstract: The growing use of large language models (LLMs) to generate forecasted inputs in discounted cash flow (DCF) valuation raises important methodological concerns related to reproducibility and transparency. DCF valuation critically depends on forecasts of key input variables. According to the principles of the International Valuation Standards (IVS), valuation conclusions must be based on a documented process that allows independent reproduction of conclusions under identical assumptions. The aim of this paper is therefore to examine whether AI-assisted forecasts of valuation-relevant inputs remain consistent when identical tasks are repeated, and to assess the implications of observed variability for the reproducibility of the valuation process. Based on an experiment involving repeated runs of an identical prompt, the study shows that differences arise not only in the wording of the responses but also in point estimates, forecast intervals, and the accompanying reasoning. The findings suggest that AI-assisted forecasting of valuation inputs cannot be considered fully reproducible in the sense required by IVS without additional control mechanisms. The paper therefore proposes the archiving of prompts, outputs, and metadata, repeated runs of identical tasks, and the explicit separation of factual inputs from expert judgement as minimum conditions for transparent and reviewable use of AI-assisted forecasts in DCF valuation.

Keywords: international valuation standards (IVS), reproducibility, transparency, AI-assisted forecasted inputs, discounted cash flow (DCF) valuation

JEL Classification codes: G17, C52

 

 

Fulltext: PDF

 

Published by: Prague University of Economics and Business, Oeconomica Publishing House

Year of publication: 2026

Online publication date: 20 May 2026

Copyright: Authors of the papers

 

ISBN 978-80-245-2587-7

ISSN 2453-6113

 

Pages 91-100

 

Back to Proceedings