The Problem of Induction: Why the Past Doesn't Prove the Future

The disclaimer on every financial product says it: past performance is not indicative of future results. Everyone reads it. Almost no one acts as if they believe it. The track record, the historical volatility, the back-test — these remain the primary inputs to most investment decisions. And the reason is understandable: if past performance tells us nothing, what does?

Nassim Taleb uses this tension to introduce one of philosophy's oldest and most practically important puzzles: the problem of induction.

Hume's Asymmetry

David Hume, writing in the 18th century, identified the fundamental logical asymmetry at the heart of empirical learning:

A thousand observations of white swans do not prove that all swans are white. A single observation of a black swan disproves the claim.

Confirming observations accumulate but never reach proof. Disconfirming observations can falsify in a single instance. The asymmetry is structural — it's built into the logic of inductive reasoning, and it can't be escaped by accumulating more confirming observations.

This matters for everything that relies on pattern recognition from past data. Your bank's risk model is trained on historical data. Your medical claim is supported by previous trials. Your investment thesis is based on what has and hasn't worked before. All of these are inductive arguments: because X was true in the past, X will be true in the future. None of them, regardless of how many confirming observations they include, are logically secure against the next observation being different.

Taleb's Turkey

The most vivid illustration in Fooled by Randomness is the turkey.

A turkey is fed every morning. For a thousand days. Each feeding is a confirming observation for the belief that the farmer is benevolent and that food arrives in the morning. The belief is well-supported. The track record is perfect. The turkey's confidence is, by any empirical standard, justified.

On day 1,001, just before Thanksgiving, the farmer arrives with a different intention.

The thousand days of feeding were not evidence that the arrangement was safe. They were evidence that the turkey had not yet observed the disconfirming event. But the disconfirming event was always in the distribution — it was the point of the arrangement, from the farmer's perspective. The turkey's inductive inference was perfectly logical and completely wrong.

The unsettling implication: you cannot tell, from inside a process, whether you are the turkey or not. The confirming observations feel identical whether or not the disconfirming event is approaching. The LTCM traders in 1997 had four years of confirming observations. The Lehman executives in early 2008 had decades of confirming observations. The turkey in October had a thousand confirming observations. The induction looked the same from inside each.

What This Does to Track Records

The practical consequence for evaluating performance is significant.

A long track record is not proof of a sound underlying process. It's evidence that a sound process has been running — or that an unsound process hasn't yet encountered the disconfirming event. From the outside, these are difficult to distinguish.

The longer the track record, the more confident most people become. But in fat-tailed environments — where the disconfirming event is rare but catastrophic — the long track record may be actively misleading. It's a thousand days of feeding. The Thanksgiving probability isn't displayed anywhere in the data.

This doesn't make track records useless. They carry information. A twenty-year track record across multiple regime changes provides better signal than a two-year track record in one favorable environment. But neither is proof. Both are inductive arguments that remain vulnerable to the disconfirming observation.

The Epistemological Implications

The induction problem changes how you should relate to your own beliefs and forecasts.

Any belief supported by past observations should carry explicit acknowledgment that it's inductive — that it's the conclusion of a "this has worked so far" argument, which is structurally different from a proof. The degree of confidence you place in the belief should be proportional to the quality and quantity of confirming evidence and the absence of disconfirming evidence — while remaining aware that the disconfirming event could arrive without warning.

This is different from radical skepticism. You can and should act on inductive inferences — there's no other way to function in an uncertain world. But acting on them with the humility appropriate to their logical status means: keeping mental track of what would falsify the belief, remaining genuinely open to updating when evidence changes, and not scaling up commitment to the belief in ways that can't survive the disconfirming event when it arrives.

The Risk Management Application

In risk management specifically, the problem of induction translates directly: never allow the model to both select the trade and size the risk.

If the model is an inductive argument from historical data — which all models are — then it will fail to see the disconfirming event, by construction. That event is the one thing the historical data doesn't contain. Using the same model to set risk limits means the risk limits fail simultaneously with the trade rationale, which is exactly the failure mode of LTCM.

The structural solution is asymmetric reliance on the model: use it for trade identification (where being wrong costs you the position), but not for risk sizing (where being wrong costs you the firm). The risk limit has to be external to the model — set by a rule that doesn't depend on the model being right.

This is Taleb's Pascal principle applied directly: use the inductive argument where you benefit from it, and don't rely on it where it can hurt you. The problem of induction isn't solved. But its damage is contained to the domain where being wrong is manageable.

For the full framework, read Living With Randomness.