| name | scenarios-horizon-and-tail-risk |
| description | Check whether a forecast's precision, form, and confidence match what is knowable at its horizon and its subject's tail behaviour; use when reviewing scenario, range, or tail-risk framing. |
| kind | skill |
| status | ready |
| provenance | {"principles":["P004","P026","P040","P047","P055","P056","P079","P081","P084","P085","P087","P089"],"claims":["C00098","C00099","C01252","C01253","C00920","C00924","C01242","C01243","C02793","C02794","C01367","C01368"],"evidence":["E00001","E00002","E00285","E00286","E00136","E00138","E00275","E00276","E00694","E00695","E00360","E00361"],"source_anchors":["6ec9d9fb28d6-c0002","e3c7c0b4e46f-c0018","e3c7c0b4e46f-c0000","e3c7c0b4e46f-c0001","e3c7c0b4e46f-c0017","a445f2941de4-c0057","5da0a790f5ae-c0002","5da0a790f5ae-c0010","5da0a790f5ae-c0009","5da0a790f5ae-c0011","5da0a790f5ae-c0012","d88ef7714f29-c0012"],"authored_from_digest":"a423039cd785971870e6f2d812633644ccd8a5e260edd1f965f85fe66a2e8109"} |
Scenarios, Horizon, and Tail Risk
Purpose
This skill reviews whether a forecast's claimed precision, form, and confidence actually match
what is knowable about its subject, at its stated horizon, given how the world it describes
really behaves. It checks whether a single-outcome estimate should instead have been built as a
branching set of scenarios, whether the forecast's confidence has been discounted for how far out
it reaches, whether the domain being forecast is quietly treated as bell-curved when it is really
fat-tailed and capable of throwing up outcomes far outside intuition, and whether the forecast's
theory of the case is one deduced storyline or a macro-theory properly gated by checkable
micro-variables and an explicit hedge. Good calibration is not only "is the number right" — it is
"is the number right for what is actually forecastable at this range, in this kind of world."
When to use
Use this skill once the estimate has passed the sibling calibration, scoring, base-rate, and
bias checks, to test its fit to horizon and its exposure to tail risk. Concretely:
- The estimate is a single confident point or narrative about a complex, high-consequence
situation, and no one has asked whether a branching set of futures would serve the
decision-maker better (P004).
- The forecast states or implies a horizon (next quarter, next year, next decade), and the
reviewer needs to judge whether the stated confidence is realistic for that distance (P040).
- The quantity being forecast is the kind that can produce extreme values far outside a
normal-curve intuition (P026).
- Someone in the discussion argues forecasting is either a solved, formulaic problem or a
worthless enterprise, and the review needs to locate the real claim between those poles (P047).
- The decision under review is a low-probability, high-payoff bet, and the reviewer must check
how much weight a small change in the probability estimate is being asked to carry (P055).
- The forecast's justification reads like an explanation of mechanism, and the reviewer must
check whether that explanatory soundness has been mistaken for predictive power (P056).
- The forecast chains two or more predicted outcomes together (a shock causing a downstream
event), and each link needs its own scrutiny (P079).
- The forecast leans toward an extreme trajectory, or toward a sharp regime break, and the
reviewer must check whether that lean is disciplined (P081, P084).
- The forecast rests on one macro-theory deducing a single outcome, or on a theoretical concept
whose predictive value has never been tested, or on a simplicity that discards real structure
(P085, P087, P089).
Procedure
Work the following checks in order; skip a check only when the input plainly does not raise it,
and say so in the finding.
-
Decide whether a single outcome is trustworthy, or whether scenarios are owed (P004). If
the situation is complex and uncertain enough that no single story should be trusted, check
whether the estimate should instead have been built as an Alternative Futures analysis: the
two most critical and uncertain drivers chosen as axes, crossed into four future worlds each
with a plausible narrative and early signposts, built so decision-makers can pressure-test
their strategies against every world. Because the technique is costly, flag its absence only
for genuinely high-consequence problems — do not demand it for a routine, low-stakes estimate.
-
Calibrate confidence to the stated horizon (P040). Predictability is not a fixed property
of a topic; it depends on what is being forecast, how far out, and under what conditions.
Check that stated or implied confidence declines as the horizon lengthens, and treat a
forecast that holds the same confidence from one year out to five-plus years out as
miscalibrated by construction — in most open systems accuracy decays toward chance by three to
five years out, and nothing much past a decade should be presented as reliable. Ask the
forecaster to scope the claim to the window it can actually support.
-
Check the shape of the distribution being forecast (P026). Before accepting a confidence
interval or a "this is too unlikely to matter" dismissal, ask whether the underlying quantity
is fat-tailed rather than bell-curved — many consequential quantities are (wealth is the
familiar case: a normal-curve intuition badly undersells how far some outcomes sit out on the
tail). If so, extreme outcomes are far likelier than that intuition suggests, and a wide
interval or a scenario branch for the tail is not overcaution. Also weigh that the present
outcome is only one draw among many once-possible worlds: a near-miss catastrophe is evidence
for taking prevention more seriously, not evidence the risk was overstated.
-
Place the claim correctly between "forecasting is a formula" and "forecasting is bunk"
(P047). If the forecast's credibility rests on a short-horizon accuracy record, flag that
this says nothing about its ability to catch rare, high-impact events — that is a different
claim needing different evidence. If the forecast is being dismissed because of one missed
rare event, check whether that event was genuinely inconceivable beforehand or merely "grey" —
imaginable, and reachable through a chain of smaller, forecastable steps that were themselves
visible. Neither extreme position should stand unexamined.
-
Weigh how much a small probability shift is being asked to carry (P055). When the decision
on the table is a low-probability, high-payoff option, choosing to proceed despite long odds
can be entirely reasonable — a large enough payoff can make even a slight chance of success
worth pursuing. Check two things: whether the estimate's precision is actually good enough to
support the fine distinction the decision turns on (it often is not), and whether the size of
the stakes is itself inflating the probability estimate — that distortion is sharpest exactly
when the decision-maker cannot yet test the estimate by acting on it.
Inputs
- The forecast or estimate itself, including its stated or implied horizon and exactly what
outcome it commits to.
- The reasoning behind it: the mechanism, theory, or causal chain it rests on, and whether that
chain links more than one prediction together.
- Whether the situation is high-complexity / high-consequence enough that a scenario-based
approach would plausibly be worth its cost, and whether one was considered.
- The stakes and payoff structure of the decision the forecast feeds, especially if it is a
low-probability, high-payoff bet.
- What is known about the historical base rate of extreme values or regime breaks in this
domain, so the fat-tail and rare-regime-change checks have something to compare against.
Output
Each finding follows the same four-part shape used across this reviewer's findings:
- Name the flaw — which of the twelve checks above it fails, and the specific way (for
example: confidence held flat across a lengthening horizon; a macro-theory deducing a single
outcome with no gating variables).
- Apply the correction — the concrete fix: switch to an Alternative Futures scenario set,
discount confidence for the horizon, widen the interval for a fat-tailed quantity, add gating
micro-variables and a hedge, or separate the explanatory claim from the predictive one.
- State the residual uncertainty — what remains unresolved even after the correction, and
the horizon or scope it is bounded to.
- Give a next step — the concrete action the forecaster or decision-maker should take next
(build the four-world matrix; re-scope the claim to the tractable window; name the
micro-variables to watch).
Order findings highest-impact first. Never end on a bare "the forecast is right/wrong" —
always close on the next step.
Anti-patterns to flag
- A single confident point forecast offered for a high-complexity, high-consequence question
with no scenario alternative considered (P004).
- Stated confidence that does not decay as the horizon lengthens, or a claim projected years past
the window the method can actually support (P040).
- A fat-tailed quantity treated as if it were bell-curved, or an extreme outcome dismissed as
impossible because it lies outside a normal-curve interval (P026).
- Either "this forecasting method is a solved formula" argued from a short-horizon track record,
or "forecasting is bunk" argued from one missed event that was actually foreseeable through
smaller forecastable steps (P047).
- A low-probability, high-payoff estimate whose precision is treated as adequate to the decision,
or whose probability looks inflated by the size of the stakes rather than by evidence (P055).
- An explanatory account of mechanism presented as if it were a tested predictive claim (P056).
- Confidence in one predicted linkage bleeding into confidence in a second, separately-uncertain
linkage (P079).
- An extreme or runaway trajectory asserted with no argument for why the usual dampening
mechanisms would fail this time (P081).
- A sharp regime change assigned a high absolute probability rather than a probability nudged up
from a low base rate (P084).
- A single outcome deduced straight from one macro-theory, with no gating micro-variables and no
hedge (P085).
- A theory or framework credited with predictive power it has never demonstrated through a
precise, nonintuitive, testable prediction (P087).
- A forecast simplified to one clean driver for elegance, in a domain where the gated
micro-variables it discarded were doing real work (P089).
References
../../references/calibration-forecasting-principles-index.md — the full principle index for this
reviewer, including the principles this skill applies (P004, P026, P040, P047, P055, P056,
P079, P081, P084, P085, P087, P089) alongside the sibling-skill principles it does not
duplicate.
Provenance
This skill body derives from P004, P026, P040, P047, P055, P056, P079, P081, P084, P085, P087,
and P089, distilled from A Tradecraft Primer, Superforecasting, Expert Political Judgment,
Perception and Misperception in International Politics, and Thinking, Fast and Slow (all
distillation-only sources; paraphrased throughout, no verbatim quotation).