belief & evidence · metaphor 5 of 100
The theory that explains every detail of your past is usually wrong about your future. A worldview can fit its training data — your childhood, your industry, your one bad decade — too well, memorizing the noise and calling it wisdom.
Every family has the burned investor: he can tell you exactly why the market fell in 2008, and again in 2020, and in the bad year that took his retirement — his system has grown a clause for each crash he has lived through, and by now it is airtight. And everyone has the heartbroken friend whose theory of what men do accounts for all four exes flawlessly: each betrayal slotted, each early warning indexed in hindsight. Ask either of them what happens next, and watch the confidence curdle into hedging.
Their theories fit. That's the problem. A life is a small, noisy sample, and a sufficiently flexible story can wrap itself around the noise — after which every new day arrives as a surprise the theory must be amended to explain away. The machine below lets you commit this exact sin with a slider, and — unlike life — it lets you watch the future arrive.
Memorizing the noise
Everything that has ever happened to you was two things at once: the tendency — how the world actually works — and the accident — the layoff that hit your division and not the next one, the person who happened to sit beside you. A finite life cannot tell you, point by point, which is which. A theory with few moving parts is forced to average over the accidents. A theory with many is free to explain them — and on a small sample that freedom is almost entirely spent explaining accidents, because accidents are most of what detail there is.
Watch the gold curve in the chart above: error on the past only falls as the theory grows more elaborate. Explaining more of what already happened is always possible — which is exactly why it is worthless as evidence that you understand anything. The cost is invisible, because it lands in the one region the fit has never seen. It stays invisible until the future shows up with data of its own.
What to try
Reveal the future, then drag the theory slider from 0 to 15. Your past rewards you the whole way — train error falls forever — while the future's error traces a U and then explodes. The truth here has degree 2. Notice how long after 2 your past keeps applauding.
Pull n down to 8. Now even a modest degree-5 theory memorizes everything, and the gap yawns open. The less life you have seen, the simpler your theory deserves to be — youth's grand systems are not wrong to be ambitious, just early.
Push σ up. In noisy domains — markets, romance, careers — there is more accident per point to memorize, so elaborate theories are punished hardest. The sage's advantage over the cynic widens exactly where life feels most unfair.
Same world, different accidents. The degree-2 curve barely moves between lives; the degree-13 curve writhes into a wholly new shape each time. A theory grown from noise inherits noise's loyalty to nothing — it isn't even the same theory twice.
The inherited prior
The fine print under the instrument admits to a whisper of ridge — a mathematical thumb on the scale that pulls wild coefficients toward zero unless the data insists. Statisticians call this a prior: a constraint adopted before your own evidence speaks, which deliberately refuses some of the fit's freedom. Inherited common sense plays the same role in a life. Proverbs, professional rules of thumb, the maddeningly boring advice of elders — these are crude, frequently wrong in the particulars, and fitted on vastly more lives than yours. Accepting them is accepting a bias in exchange for stability.
The lone theorist who deduces everything from his own four data points owes the ancestors nothing — and carries maximal variance. His system is original in precisely the way the degree-13 curve is original: shaped by his accidents, transferable to no one, due for revision at the next surprise. It is an argument about error budgets: when your sample is small and your domain noisy, borrowing someone else's smoothing is not cowardice. It is statistics.
Field diagnostics
You can score a theory without waiting a decade, because overfit stories share a signature. They explain everything retrospectively and predict almost nothing: fluent about every ex, silent about the next date. They need a new clause after each surprise — the epicycle, the special case, the "this only proves how deep it goes" — each amendment fitting the last data point and forecasting none of the next, a degenerating research programme in miniature, as Lakatos (sharpening Popper) diagnosed in science. And they are proud of their own detail, offering the sheer number of things accounted for as credentials — when past a point, that number is the pathology. The remedy is cross-validation, practiced on yourself: write down what your theory expects before the future arrives, and score it only on that.
The mapping
| Mathematics | Life |
|---|---|
| data points | What has happened to you — the events of one finite life, one per dot. |
| the true curve | How the world actually tends — real, simple-ish, and never directly visible. |
| noise ε | The accidents that happened to happen: real events carrying no lesson. |
| degree d | The elaborateness of your explanation — its clauses, exceptions, and villains. |
| train error | How well the story fits your past. Falls forever; flatters always; proves nothing. |
| test error | How the story meets your future — the only score that was ever about the world. |
| the U-curve | Why both naïveté and cynicism predict badly: too few clauses miss the shape, too many memorize the pain. |
Where the metaphor tears
The instrument tempts you to worship the low degree, but underfitting is the equal and opposite sin — bias is also an error term. The simpleton's flat line is stable, portable, and wrong about the entire shape of things, and some domains genuinely are complicated: their true curve has more bends than a proverb can hold. The lesson is matching elaborateness to evidence, never minimizing it on principle.
In the sandbox, the next decade is drawn from the same world as the past — the kindest possible test. Lives are not like that: regimes change, industries die, you yourself become a different sampler of experience. When the world moves, a theory can fail while fitted perfectly well — not variance but obsolescence. That failure mode is systematically worse than anything the overfitting lens prices in, and no degree slider repairs it.
People fit theories to their pasts partly for meaning, not forecasting. The narrative that makes a painful decade coherent can be predictively overfit and still psychologically load-bearing — it may be what lets its owner get up in the morning. This lens can price such a story's forecasting record precisely; what the story holds up, it cannot weigh at all.