Regret Minimization: A Framing Device, Not a Scoring Algorithm

“Regret minimization” means two different things that get used almost interchangeably, and it is worth separating them before using either one. The first is a qualitative mental exercise: imagine looking back on a choice from a distance — years from now, or simply after the dust has settled — and ask which option you would regret more, the one you took or the one you passed up. It requires no numbers, no scenarios, no payoffs; it is a framing device for surfacing an intuition. The second is a formal decision rule from statistical decision theory, sometimes called the minimax-regret criterion, which is what the regret minimization scorer on this site actually computes: a specific arithmetic procedure over explicit options, explicit scenarios, and explicit payoff numbers. The two are related in spirit but are not the same tool, and treating the second as a precise, algorithmic version of the first — rather than a narrower formalization that only handles part of what the first is getting at — is the most common way this framework gets misused.

What the formal tool actually computes

The minimax-regret criterion is built for a specific situation: you can list your options, you can list the possible future scenarios that might unfold, and you can estimate how well each option would pay off in each scenario — but you cannot confidently put probabilities on which scenario will actually happen. Given that setup, the tool computes, for every scenario, the best payoff any option achieves in it. Then, for each option, it works out that option’s regret in each scenario — how far below the scenario’s best payoff that option actually fell. An option’s overall regret score is its worst regret across all the scenarios, not its average, and the recommended option is whichever one has the smallest worst-case regret: the option you would be least sorry about, no matter which scenario turns out to be the one that actually happens.

A worked example: two commute routes

Suppose you are choosing between two ways to get to a regular commitment: a highway route that is fast when traffic is light, and a set of surface streets that is slower on a good day but far less exposed to a highway backup. You define three traffic scenarios — light, normal, and heavy or an accident — and estimate each route’s time saved, in minutes, relative to a familiar 40-minute baseline, in each one:

• Highway — light traffic: saves 15 minutes; normal traffic: saves 5 minutes; heavy traffic or an accident: costs an extra 20 minutes.
• Surface streets — light traffic: saves 5 minutes; normal traffic: saves 8 minutes; heavy traffic: costs an extra 5 minutes.

Running these through the scorer, the best payoff available in each scenario is 15 minutes (light), 8 minutes (normal), and −5 minutes (heavy) — the surface streets actually win outright in normal and heavy conditions, with the highway only pulling ahead when traffic is light. The highway’s regret in each scenario is then 0, 3, and 15 minutes respectively, for a worst-case regret of 15. Surface streets regret 10, 0, and 0 minutes, for a worst-case regret of 10. Even though the highway has the single best best-case outcome of anything in the table (15 minutes saved on a light-traffic day), the tool recommends surface streets, because its worst case is meaningfully less painful than the highway’s worst case — a 35-minute swing on a bad highway day, against a comparatively mild 5-minute cost on a bad surface-streets day.

What that arithmetic is deliberately not doing

Notice what minimax regret does not ask for anywhere in that calculation: it never asks how likely heavy traffic actually is. That is the whole point of the criterion — it is built for exactly the situation where you do not trust your own probability estimates enough to weight the scenarios by them, so it protects you against the worst case regardless of how likely that worst case seems. That is also its most important limitation, and one worth being explicit about rather than letting the confident-looking output paper over it: if heavy traffic is genuinely a once-a-year event, minimax regret still treats avoiding its regret as the entire deciding factor, potentially leaving a route that is meaningfully better nearly all the time on the table because of a scenario you will rarely actually face. When you do have workable probability estimates for the scenarios, an expected value calculation that weights each scenario by how likely it actually is will typically serve you better than a criterion that is deliberately blind to likelihood. Minimax regret earns its keep specifically in the gap where probabilities are unavailable or untrustworthy, not as a general-purpose replacement for them.

The larger gap: this still isn't "regret"

Even setting probability aside, the scorer’s output is not a measurement of regret in the psychological sense — it is a measurement of a payoff shortfall, which is being used as a stand-in for regret because it is something a formula can actually compute. Real anticipated regret is shaped by several things a payoff table cannot represent. People generally feel differently about a bad outcome that followed from something they actively chose to do versus one that followed from inaction or the status quo, even when the payoff numbers are identical — the psychology literature broadly describes this as an asymmetry between action and inaction, though the size and conditions of that asymmetry are debated and this piece will not attach a number to it. Regret is also shaped by visibility and blame: an outcome other people will see and judge you for tends to be anticipated more painfully than an equally bad outcome no one else will ever know about, again independent of the raw payoff. None of that shows up in a scenario table of minutes saved or dollars gained, because a scenario table only has room for the payoff itself, stripped of the story around how you got there. The scorer is measuring something real and useful — worst-case payoff shortfall across an explicit set of scenarios — but calling that measurement “regret” is a convenient shorthand, not a literal claim that the number captures the feeling.

The qualitative version, and where it actually shines

The framing device that regret minimization is best known for — imagining yourself looking back on a choice from a distance and asking which option you would regret more — does not require any of the formal machinery above, and for a lot of decisions it should not be forced into it. It works precisely because it is loose: it lets you draw on your whole sense of who you are and what you care about, not just the handful of numeric criteria that fit cleanly into a table. That looseness is a strength for the kind of decision where the real stakes are hard to reduce to comparable payoffs across a short list of scenarios in the first place, and a weakness the moment you try to dress it up as a precise calculation it was never built to be. Used honestly, it is a prompt for reflection — a way to step outside the immediate pressure of a choice and ask a longer-range question — not a verdict, and not a substitute for actually thinking through the specifics of your situation, or for talking it through with people who understand it better than a framing device ever could.

A boundary worth naming

Because the qualitative framing feels so natural to apply to any decision at all, it is tempting to reach for it on choices that really deserve a different kind of attention — a major career move, the end of a relationship, a decision genuinely entangled with someone’s wellbeing. Imagining your future self’s regret can be a useful piece of how you think about those situations, but it is a prompt for your own reflection, not a verdict, and it is not a substitute for the kind of conversation those situations usually deserve with people who actually know the specifics — a mentor, a partner, a professional, depending on what is actually at stake. A framing device that helps you organize your own thinking before that conversation is doing exactly its job; one that gets treated as having already settled the question in place of it is being asked to do more than it honestly can.

Choosing between the two versions

Reach for the formal scorer when you can genuinely name a short list of distinct scenarios and assign each option a comparable, honest payoff number in each one — the commute example above, choosing between suppliers under different demand scenarios, deciding how to staff a project against different possible timelines. Reach for the qualitative “looking back from a distance” framing instead when the thing you are actually weighing resists being reduced to a short list of scenarios and comparable numbers at all, and treat its output as a prompt to examine your own reasoning more closely, not as an answer in itself. The two are not competitors; they are simply built for different shapes of uncertainty, and the mistake to avoid is applying the confident, mechanical output of one where only the honest, open-ended reflection of the other actually fits the decision in front of you.