A real option is the right, but not the obligation, to take an action on a physical asset later: delay a plant retrofit, abandon a building phase, switch a component. Its value comes from waiting to see how the world turns out. Textbook valuations assume the manager knows the exact probabilities with which prices move, usually through a smooth random-walk model called geometric Brownian motion. In the projects where options matter most, such as brownfield redevelopment or carbon capture retrofits, those probabilities come from a handful of experts who give ranges, not numbers. This note proposes to value such options on a small, finite tree of future states in which each transition is described by a set of plausible probability tables rather than one table. The work then asks a question the literature has not posed. Does the way those stage-by-stage ranges are stitched into a joint model change whether the manager should invest now or wait?
The central observation is that the stitching choice is not a technicality. If the ranges are treated as re-drawn independently at every period, which is the standard "rectangular" construction that makes backward induction work, then observing today's price says nothing about tomorrow's probabilities. Ambiguity can then only lower the option's value. If there is instead one true but unknown transition table, fixed across time, then waiting also reveals which table is true, and the option earns an extra learning premium. The same expert ranges can therefore make an ambiguity-aware manager either more eager or more reluctant to invest. We propose to measure that effect, to identify when it flips decisions, and to test which construction gives better realized outcomes in worlds where the true model is known to the experimenter.
Core Claim
On a finite state tree with interval-valued expert beliefs, the value of a deferral option splits into two parts. The first is an aleatory part, the familiar value of waiting out randomness. The second is an epistemic part, whose sign depends on how stage-wise beliefs are assembled into a joint model. Under rectangular assembly the epistemic part is a pure penalty. Under fixed-model assembly with likelihood-based learning it can be a premium. We expect the invest/wait decision to differ between the two over a non-trivial region of parameter space, and we expect the learning construction to achieve lower realized regret when the true kernel is actually stable. This gives a concrete, testable form to a tension the research file flags but does not resolve: in option theory "uncertainty is an asset," while in imprecise-probability theory "ambiguity is under-determination."
This claim does not say that ambiguity-aware real options are new. Maxmin-style ambiguity models of investment timing exist in the economics literature, though none appear in the material available to us, so we make no claim over them. What we claim is new is treating the assembly step, meaning which joint model is built from stage-wise beliefs, as the object of study. We also claim novelty in linking two results from the belief-representation literature to a concrete exercise decision, and in testing the constructions against a known ground truth.
How the Two Literatures Meet
Three results from the belief-representation literature supply the problem, and each plays a different role.
The first result motivates the problem. The Total Belief Theorem generalizes the law of total probability to Dempster–Shafer belief functions. These functions assign weight to sets of outcomes rather than single outcomes. The theorem shows that a prior plus conditional beliefs pins down a unique joint belief only when the prior is an ordinary probability distribution, and that otherwise infinitely many joints fit. In an option tree, "prior" and "conditional" correspond to today's beliefs and the stage-wise transition beliefs. Nobody has asked whether the different admissible joints recommend different exercise times. We ask exactly that.
The second result supplies both the learning mechanism and a warning. Weighted Sets of Probabilities and Minimax Weighted Expected Regret attaches a plausibility weight to each candidate probability measure and updates the weights by likelihood, $\alpha_k \propto \alpha_k P_k(E)$. It pairs this with a regret-minimizing choice rule. The same paper notes that such agents can be dynamically inconsistent: a plan that is optimal today may be abandoned tomorrow. That warning constrains our design. We cannot simply run backward induction on the regret criterion. We must compare a naive planner with a sophisticated one that anticipates its future choices.
The third result acts as a gatekeeper on representation. Conditions for the Existence of Belief Functions Corresponding to Intervals of Belief gives necessary and sufficient conditions for expert intervals on atoms to be representable as a belief function. It reports that in a three-event case only about one-fifth of probabilistically consistent intervals qualify. This finding cuts against making Dempster–Shafer the core representation. We therefore use general sets of probabilities, called credal sets, as the working object. Belief functions are a special case, and the Lemmer–Kyburg test is a diagnostic telling us when the Total Belief Theorem's analysis applies directly. Coherence checking follows From imprecise probability assessments to conditional probabilities with quasi additive classes of conditioning events, which gives constructive procedures for turning interval assessments that avoid sure loss into coherent conditional probabilities.
On the real options side, the closest existing bridge is Brownfield Redevelopment Evaluation Based on Fuzzy Real Options. It turns five experts' interval estimates into fuzzy numbers and feeds them into the continuous-time Black–Scholes formula to value a two-year option to wait. It never checks whether the intervals are coherent, it keeps continuous dynamics, and it has no notion of learning. Minimax Regret Optimisation for Robust Planning in Uncertain Markov Decision Processes is the closest method. It derives a regret Bellman equation for finite planning problems whose transitions lie in uncertainty sets, and it handles dependence across time with multi-step macro-actions. Its "options" are planning macro-actions, not economic exercise rights. Its uncertainty sets are flat, with no weights and no learning. Halpern's A Counterexample to Theorems of Cox and Fine serves only as motivation: on finite domains the classic argument that rational belief must be probability does not go through, so insisting on a single precise table is a modeling choice, not a requirement. We do not claim this breaks option pricing.
Research Program
We build a finite-state option valuer with three assembly regimes that share the same expert inputs. The inputs are a lattice of states $s$ (for example, discretized carbon or oil prices), an exercise payoff $g(s)$, a discount rate $r$, and for each state a credal set $C(s)$ of transition tables, elicited as intervals and repaired to coherence. The rectangular regime values the option by
$$V_t(s)=\max\Big\{g(s),\ \tfrac{1}{1+r}\min_{P\in C(s)}\textstyle\sum_{s'}P(s'\mid s)\,V_{t+1}(s')\Big\},$$
which is dynamically consistent and solvable by one small linear program per node.
The fixed-model regime instead posits a finite family of candidate kernels $\{P_k\}$, obtained by sampling or taking vertices of the credal sets. It carries a weight vector $\alpha$ as part of the state:
$$V_t(s,\alpha)=\max\Big\{g(s),\ \tfrac{1}{1+r}\,\Phi_\alpha\big[V_{t+1}(s',\alpha')\big]\Big\},\qquad \alpha'_k\propto\alpha_k P_k(s'\mid s).$$
Here $\Phi_\alpha$ is the aggregation across candidate kernels. We run it in three versions: weighted expectation, which is Bayesian with a hyperprior; weighted maxmin; and MWER, the minimax weighted expected regret rule. For MWER we compute both the naive plan and a sophisticated plan in which each future self's choice is taken as given. The gap between the two directly measures dynamic inconsistency.
The third regime enumerates admissible joint belief functions via the Total Belief Theorem's ballooning-extension construction. It applies only when the Lemmer–Kyburg conditions hold. It yields the full range of option values and exercise policies consistent with the prior and conditionals.
Two quantities come out of this. The first is the learning premium: fixed-model value minus rectangular value, with the same aggregation rule so that only assembly differs. The second is the exercise-indeterminacy region: the set of states and times where admissible models disagree on invest versus wait. For small trees the weight-augmented state space can be solved exactly by discretizing $\alpha$. For larger trees we would use least-squares Monte Carlo regression on $(s,\alpha)$, as in the CCUS literature.
Baselines
The naive reference point is static net present value: commit now or never, using midpoint expert estimates. A Systematic Literature Review on the Traditional NPV Model and Its Improved Versions documents why this ignores flexibility. On the brownfield case, Dai, Sun and Guo report a static NPV of −18,212 Yuan, which would reject the project. Their fuzzy option approach gives a waiting-option value of 11,866 Yuan and a project value of 30,078 Yuan, which accepts it. We recompute static NPV on every benchmark with the same inputs and discount rate.
The second baseline is precise-probability real options on the same finite lattice, using one representative table per state. We use either the interval midpoint or the pignistic transform, which spreads each belief mass evenly over the outcomes it covers. This is what a careful practitioner would do. It shows how much of any difference comes from imprecision itself rather than from the discretization.
The strongest existing method for the robust comparison is rectangular regret or maxmin planning in the style of Minimax Regret Optimisation for Robust Planning in Uncertain Markov Decision Processes, with exercise-or-continue as the action set. That paper reports its regret method beating maxmin and averaged-model baselines on normalized maximum regret in its medical, disaster-rescue and glider domains. The file's summary gives no numeric values, so there is no reported result found on Lacuna for the magnitude of that gap. All robust methods receive identical credal sets, identical lattices, and identical compute: exact solution on small trees, and the same number of regression paths on large ones.
We also run a fuzzy Black–Scholes reimplementation following Dai, Sun and Guo, as the closest published expert-interval option method.
The key ablation holds the aggregation rule fixed and changes only assembly, rectangular versus fixed-model. It also runs fixed-model with learning switched off, freezing $\alpha$, which isolates the learning channel. Beating the baselines means lower realized regret against a hindsight oracle, averaged over synthetic worlds whose true kernel lies in the credal set. The target is a margin that exceeds the variation across random seeds and across hyperprior choices for the Bayesian variant.
Why not just…?
Why not just use a single precise table from the expert midpoints? It is the cheapest option and will often agree with our method. The proposal predicts where it will not agree: inside the exercise-indeterminacy region, which will be measured directly. If that region is empty or negligible on realistic parameters, midpoint valuation is adequate, and that would be a useful negative result. The midpoint approach cannot say when its own decision is fragile.
Why not just put a Bayesian hyperprior over the candidate kernels? This is the most serious threat, because a hierarchical Bayesian also learns and also earns a learning premium. Our weighted-expectation variant is that model, so the comparison is built in. The distinguishing experiment sweeps the hyperprior (uniform over vertices, uniform over the credal volume, concentrated near midpoints) and records how often the invest/wait decision flips. If no reasonable hyperprior changes decisions, the Bayesian model wins on simplicity and we will say so. The imprecise regimes are justified only where the Bayesian answer is hostage to an arbitrary hyperprior, which is what the finite-domain results warn can happen when evidence is thin.
Why not just run robust maxmin backward induction on the credal lattice? It is tractable and dynamically consistent. But the rectangular recursion it relies on makes learning impossible by construction, so it can only penalize ambiguity. Our central prediction is that it will systematically undervalue waiting, and so recommend premature investment or rejection, when the true kernel is stable. The regret-versus-oracle experiment tests this directly. Maxmin should do comparatively well only in synthetic worlds whose kernel really is re-drawn each period, and we include such worlds so that the test can go either way.
Why not just use fuzzy real options as published? The fuzzy approach keeps continuous Black–Scholes dynamics and never checks the expert intervals for coherence. We test it on the same inputs. We also report how many realistic interval elicitations fail coherence or belief-function representability, where Lemmer and Kyburg's roughly one-fifth figure for three events gives the prior expectation. For those inputs, fuzzy valuation has no well-defined underlying probability model.
Proposed Experiments
The decisive test is synthetic. We generate finite deferral and abandonment problems on lattices of 5 to 30 price states over 5 to 15 periods. The credal sets are built from interval widths that we vary systematically. The true kernel is either fixed and hidden (the stable world) or re-drawn each period within the set (the rectangular world). Every method sees the same credal sets. We score realized discounted payoff and regret against an oracle that knows the true kernel. Our hypothesis passes if fixed-model learning beats rectangular maxmin and regret planning in stable worlds, and does not lose badly in rectangular worlds. It fails if the learning premium is within seed noise or if the Bayesian hyperprior variant matches it everywhere.
The applied tests use the three cases in the file. For the brownfield project we have investment of 432,000 Yuan, $r$ of 5%, a two-year horizon, and five experts. The file does not report the raw expert intervals, so we reconstruct plausible intervals around the reported expected DCF of 413,788 Yuan. We ask whether acceptance remains robust across all admissible models or falls inside the indeterminacy region. For the CCUS retrofit, Strategic Decision-Making for Carbon Capture, Utilization, and Storage in Coal-Fired Power Plants reports a deferred-option value of 462 million CNY with optimal investment in year 10 under the full market setting. We discretize its setting and test whether year-10 timing survives interval-valued price drift. For the industrial-symbiosis case in Investment Decisions of CCUS Projects in China Considering the Supply–Demand Relationship of CO2 we use the reported inputs: carbon price 7.8 USD/ton with volatility 0.03 and growth 0.025, and oil price 80.9 USD/barrel with volatility 0.21 and growth 0.04. We replace the point growth rates with intervals and measure how far the reported 75.72 USD/barrel immediate-investment threshold moves under each assembly regime.
Finally, for the subset of elicited intervals that pass the Lemmer–Kyburg conditions, we enumerate admissible joint belief functions. We report the fraction of cases where two admissible joints recommend opposite exercise times. This is the first direct measurement of whether the Total Belief Theorem's non-uniqueness has decision consequences. A near-zero fraction would show that the non-uniqueness is harmless for deferral decisions, and that too is worth knowing.
References
- The Total Belief Theorem
- Weighted Sets of Probabilities and MinimaxWeighted Expected Regret: New Approaches for Representing Uncertainty and Making Decisions
- Conditions for the Existence of Belief Functions Corresponding to Intervals of Belief
- From imprecise probability assessments to conditional probabilities with quasi additive classes of conditioning events
- Brownfield Redevelopment Evaluation Based on Fuzzy Real Options
- Minimax Regret Optimisation for Robust Planning in Uncertain Markov Decision Processes
- A Counterexample to Theorems of Cox and Fine
- A Systematic Literature Review on the Traditional NPV Model and Its Improved Versions
- Strategic Decision-Making for Carbon Capture, Utilization, and Storage in Coal-Fired Power Plants: The Roles of Pollution Right Trading and Environmental Benefits
- Investment Decisions of CCUS Projects in China Considering the Supply–Demand Relationship of CO2 from the Industry Symbiosis Perspective