Put two choices side by side. For each one, list what could happen, how likely it is, and how much it would sting. Then mark whether the regret would be about something you did or something you didn't do, because the two fade at different speeds.
how long regret sticks around
ρ is the share of last year's regret you still carry this year. At 0.4, most of it is gone in two or three years. At 0.9 it barely moves. Research on regret (Gilovich and Medvec, 1995) found that what people regret over a lifetime is mostly things they didn't do, which is why the second slider starts higher.
use your own history
Demo data is on, so this uses the made-up person's history.
add self-deception
What you feel is not always what happened. Here the bias wanders over time around an average, so a person who tends to inflate their mistakes (positive) or wave them off (negative) gets a band instead of a single line.
add triggers
A song, a photo or a street can bring an old regret back at full strength. Triggers arrive at random, each one adds a spike that then fades like everything else, and the band shows the spread across 600 simulated lives.
Curves show accumulated regret: the weight of the choice if you keep making it every year, which is the same number as the total weight of one choice that fades by ρ each year.
how sure can you be?
Each bar is a share of the 2,000 runs, placed by how much lighter one choice ended up than the other.
what the answer hinges on
Each bar swings one number from the low to the high end of its range while everything else stays put, and shows which way that pushes the comparison. The dotted line is where things stand now. Numbers marked ● can flip the answer on their own.
Most choices aren't one-offs. What happens next opens some doors and closes others, and you get to choose again. Lay that out here as a tree: ■ choices you make and ● things you wait to find out. The app works backward from the end and picks, at every choice, the option that carries the least expected regret.
Pink is the best plan. Faded branches are options the plan passes on. Each number is the expected regret from that point on.
edit the tree
Every decision you save from the simulator comes with a guess about how much it will sting. Come back in three months and say how it actually feels. Over time this shows whether your guesses run high or low, and by how much.
your forecasts vs what happened
Dots on the dashed line were guessed right. Below it, it hurt less than you expected. Pink dots are choices to act, dark ones are choices not to.
A regret pays off only if the lesson gets used. Log a mistake and what it taught you, then check back later: say how much it still stings and whether you applied the lesson. That gives you a measured β for each kind of mistake, and your own ρ, how fast your regret actually fades.
how fast your regret fades
by kind of mistake
β: each check-in counts as a trial (applied it = 1, partly = 0.5, did it again = 0). It starts at 0.5 with no evidence and moves as check-ins come in. The bar shows the range β plausibly sits in (80% interval). ρ is fitted from the "still stings" ratings and pulled toward the overall figure when there are only a few.
expected regret
Each option has outcomes with probability pi and regret Ri. The regret you expect from one round is
R̄ = Σ pi Ri
regret that carries over
A share ρ of last period's regret comes with you into this one:
Et = R̄ + ρ Et−1
which solves to
Et = R̄ (1 − ρt) / (1 − ρ) + ρt E0
and settles, for ρ below 1, at
E∞ = R̄ / (1 − ρ)
With R̄ = 2, ρ = 0.6 settles at 5 and ρ = 0.95 settles at 40. The persistence term does most of the work: going from 0.6 to 0.95 multiplies the long-run weight by eight while the regret itself is unchanged. At ρ = 1 the balance grows by R̄ every period and never settles.
The simulator splits R̄ into regret over things you did and regret over things you didn't, and gives each its own ρ. Since the recurrence is linear, the two parts add.
the insurance trap
Choosing the option with the lowest regret right now tends to favour doing nothing, because inaction rarely stings much at first. But omission regret fades slowly. A bold option with a large but short-lived regret can end up lighter than a safe option with a small regret that never leaves. The simulator marks the year where the curves cross.
uncertain inputs
Every chance and sting gets a range. The "how sure" dial sets its width (up to ±30 points on a chance and ±4 on a sting), or you can set the ends yourself. Both ρ values get ±0.1, since nobody knows their own persistence exactly. Each run draws every number from a triangular distribution over its range, peaked at the value you typed, rescales the chances to sum to 1, and computes ET for both options. Across 2,000 runs that gives the share in which each option comes out lighter.
The expected value of perfect information is how much regret you'd save, on average, if you could see the true numbers before choosing:
EVPI = min(E[A], E[B]) − E[min(A, B)]
For a single number θ, the same idea with only θ revealed (EVPPI) is estimated by sorting the runs by θ into 20 equal bins and choosing the better option within each bin:
EVPPI(θ) ≈ min(E[A], E[B]) − Σbins wk min(E[A | k], E[B | k])
The sensitivity chart is one-way: each number, including both ρ values (±0.1), moves alone across its range while the rest stay at their typed values.
perceived regret
What you feel is the objective regret plus a bias that drifts around an average:
Rpert = Robjt + bt
bt = μb + φ (bt−1 − μb) + εt, εt ~ N(0, σε2)
E[Rper∞] = (R̄ + μb) / (1 − ρ)
You can't estimate μb, φ and σε for yourself without an outside view of what actually happened, so the simulator treats them as dials and shows the 10th to 90th percentile of 600 simulated paths. The bias carries over at the option's average ρ, weighted by how much of its regret is action vs omission.
triggers
Triggers are a compound Poisson process added to the same recurrence:
Et = R̄ + Jt + ρ Et−1, Jt = Σk ≤ Nt Xk, Nt ~ Poisson(λ), Xk ~ Exp(mean s R̄)
On average this only scales the long-run level by (1 + λs). The spread is what changes: a few unlucky years leave a much heavier tail, which the 90th percentile in the table shows.
decision trees
A tree has choices (■) and chance events (●). A chance event resolves some years after the moment before it, and each outcome's regret r lands then and is carried with its kind's ρ until year T:
burden(r, ρ, τ) = r (1 − ρτ+1) / (1 − ρ), τ = T − time it lands
The value of a chance event is the probability-weighted sum of each outcome's burden plus the value of whatever follows it. The value of a choice is the smallest value among its options. Working this backward from the leaves gives the plan with the least expected regret, the same rollback used in decision analysis.
forecast calibration
Each saved decision has a predicted sting and, later, an actual one. The bias is the average of actual − predicted, pulled toward zero as if you'd also made k = 2 perfect forecasts:
bias = n · d̄ / (n + k), 80% range: bias ± 1.28 · s / √(n + k)
With five or more rated decisions the app also fits actual = a + b · predicted by least squares. A slope below 1 means your forecasts overshoot in both directions.
your own ρ
Each lesson-log check-in records how much the mistake still stings. If regret fades as R0 ρt, then
ln((st + ½) / (R0 + ½)) = t ln ρ
(the ½ keeps a sting of zero from blowing up the log). ln ρ is fitted by least squares through the origin. With little data that's noisy, so the estimates are partially pooled: each kind of regret (did, didn't) is pulled toward the app's default, and each kind of mistake toward its kind, with a prior worth 0.25 year² of evidence:
ln ρ̂g = (Σ x y + κ ln ρparent) / (Σ x² + κ)
whether the regret paid for itself
ΔU = −R0 + βL
L is what the lesson is worth and β is how reliably you apply it. The lesson log estimates β per kind of mistake with a Beta(1, 1) prior updated by your check-ins, and uses the posterior mean in ΔU.
the levers
Lower the chance of the bad outcomes (pi). Shrink how much they hurt by treating them as lessons (Ri). Carry less of it forward (ρ). Get more out of what it taught you (β and L).