Superforecasters rarely open a hard question with a single gut number. They decompose: start with how often this class of event happens, then list what makes this instance different. That is how an outside-view prior becomes auditable instead of a vibe you paste from a headline.
Decomposition is the bridge between storytelling and probability. You are approximating P(event) as a base rate plus adjustments you can defend, challenge, and score when the world resolves.
The two layers
The base rate asks: in the reference class I trust, how often does YES happen? Historical tables, long-run market behavior on similar contracts, and academic base rates belong here. The specifics ask: what moves this case away from the class? Polls, filings, injury reports, whip counts, and fresh micro-data belong here—each tied to a direction and rough magnitude.
Bayesian language fits naturally: the base rate is your prior; each specific piece of evidence is a likelihood-shaped nudge. The chapter on Bayesian updating makes that formal; decomposition is how you choose a defensible prior before the news flood begins.
Think of Fermi estimation in physics: break an impossible total into pieces you can estimate separately. Superforecasting does the same for probability. The whole is only as good as the pieces—and the discipline to merge correlated pieces before you count them twice.
Why monolithic intuition fails
Saying "feels like sixty-five percent" hides your reasoning. Saying "base fifty-two percent for incumbents in this state band, plus six for approval, plus four for fundraising, minus eight for open investigation, plus five for head-to-head polls, compose to about fifty-nine percent" exposes which row would flip you below market if it changed. When resolution wording is fuzzy, you widen uncertainty on a row instead of pretending minus eight percent is precise.
Correlated specifics are a classic trap. Polls, sentiment, and "momentum" often measure the same latent factor. Merge them before you add adjustments, or you will double-count enthusiasm as independent evidence. The fallacies chapter on conjunction errors is the formal warning; decomposition is the practical fix.
Worked example: incumbent re-election
Contract: "Senator Smith wins the general election" YES.
Start with a reference class: incumbents in competitive states over the last twenty years win roughly fifty-two percent of the time. That is your outside anchor. Then specifics: approval above fifty percent historically adds perhaps six points in your backtest; a third-quarter fundraising lead adds four; an open ethics investigation subtracts eight if resolution text clearly covers it; a head-to-head polling average adds five. Compose: 52 + 6 + 4 − 8 + 5 = fifty-nine percent, rounded to a trading band of fifty-eight to sixty.
If market YES is fifty-five cents, you have a candidate edge—only if fees and spread allow execution. Decomposition also tells you what to watch: if the investigation is dropped without a resolution change, maybe only four points return, not eight. If resolution criteria shift mid-campaign, you reset the table instead of nudging one cell.
Worked example: rate decision
Contract: "Fed cuts twenty-five basis points at March FOMC" YES.
Outside view: in the last thirty years, when inflation is above three percent year-over-year, March cuts from a similar futures path might be only eighteen percent. Specifics: dot plot median adds twelve points; futures-implied OIS adds fifteen; a dovish chair speech adds eight. Raw sum lands near fifty-three percent.
Experienced forecasters shrink aggressive compositions toward market or survey consensus when factors are noisy—landing near fifty percent instead of insisting on fifty-three with false precision. Conjunction-style errors appear when traders treat correlated positives as independent bonuses; decomposition forces you to see the pile.
Macro markets punish false precision twice: once in Brier when you are wrong, once in P&L when you pay wide spreads for a two-cent phantom edge.
Adjustment discipline
On the twenty-to-eighty percent scale, small additive point adjustments are readable. Large shifts are often better handled by multiplying odds so you do not slam past zero or one hundred percent. If two specifics share a cause, merge them once.
Maintain a personal base-rate library the way you maintain market dossiers: Supreme Court cert grants near twelve percent for a posture; house bills passing chamber under divided government near eight percent; FDA approvals on PDUFA dates near sixty-eight percent in your last review. When you need a prior under time pressure, you look up a row instead of panic-searching.
A library is living documentation. After each resolution, ask whether the reference class you chose was right. Wrong class is a different mistake than wrong adjustment.
Sensitivity before size
For any trade above probe size, sketch low, mid, and high for the controversial rows. If scandal impact might be minus four, minus eight, or minus fifteen points, only the optimistic scandal case should still clear costs versus executable price. If only the mid case works, you are on a watchlist, not a conviction ticket.
Sensitivity is how decomposition talks to position sizing. Kelly-style frameworks assume you trust the point estimate; honest decomposition assumes you trust a range and trade only when the range clears the book.
Multi-outcome and categorical markets
Multi-outcome markets decompose per candidate, not one story for "the race." Each leg gets a base from historical "wins from this position" plus specifics; renormalize so probabilities sum to roughly one hundred percent before you compare each leg to venue prices.
A nomination race with five serious candidates is often more honest as five contracts than as five separate binary markets that might contradict each other. Decomposition exposes which leg is rich, not merely that "the race feels close."
Linking decomposition to mispricing hunts
When you run a structured mispricing audit, the prior step should cite your base-rate row. The read step lists specifics with evidence tier. The compute step uses composed belief, not story peak. The document step stores the table, not one number.
If you cannot write the table, you have a narrative—not a trade thesis.
When decomposition says pass
If composed belief sits within a couple of cents of consensus, you usually wait. If every sensitivity row kills expected value, the audit should stop. If two reference classes fight—"incumbent base rate" versus "open seat base rate"—run the outside–inside blend discipline before clicking.
Passing is a feature. Decomposition often ends at "no edge" faster than intuition, which saves fees and calibration damage.
Worked example: multi-outcome nomination
Five candidates, one winner. Candidate A might carry a twenty-eight percent base from historical "wins from this position," plus four points for endorsements, landing near thirty-two percent. Candidate C might start at eighteen percent plus six for polling momentum, near twenty-four percent. Sum raw legs, renormalize from slight overshoot to one hundred percent, then compare each leg's ask.
You might find the market prices Candidate C rich while Candidate D is cheap—insight decomposition gives you, not "the race is close."
Common misconceptions
Decomposition is not the same as listing reasons you like a trade. Every row needs a direction, magnitude, and source. It is also not an excuse to add until you reach your desired number—if rows sum to eighty-five percent and market is fifty-five cents, check double-counting before declaring genius.
What comes next
Decomposition separates reference-class statistics from case-specific causality. The next chapter tackles the deeper tension: when to trust the outside view, when to let the inside view move you, and how not to blend them incoherently.
Key ideas to carry forward
Write the base rate first. List specifics with signs and sizes. Merge correlated rows. Sensitivity gates size. Tables beat vibes.
Decomposition is the antidote to "I like this candidate." Liking is not forbidden—but it must appear as a labeled row with a magnitude you will score when the election resolves.
The chapters ahead on outside–inside view and Bayes assume you can produce this table on demand. Make it a habit before you open the order ticket.
Next: The Outside View vs. The Inside View