Why Base Rates Anchor Good Judgment
Disagreeing with a model is easy; disagreeing with it responsibly is harder. The base rate represents everything the model knows about similar situations before your particular case arrives. When you override it, you are implicitly claiming your case differs in a way the model cannot see. Sometimes that claim is justified—models built on historical data miss regime changes, novel circumstances, and local knowledge. But the burden of proof sits with you, not the model. The discipline is to articulate exactly why this instance falls outside the reference class, and how much you would have expected the outcome to differ if your reasoning is right.
Also worth reading: When Should You Trust an Expert Under Uncertainty—and When Should You Ask for the Mechanism? · Essays on Judgment, Technology, and Uncertainty: Can Better Decisions Survive Automation?
A practical approach: state the base rate first, then state your adjustment, then quantify the gap. If you think a startup's survival odds are better than the fifteen percent base rate, say what evidence moves the number and by how much. Vague conviction—"this one feels different"—is not adjustment, it is substitution. The goal is not deference to models but a visible ledger showing where your judgment departs from the statistical prior and why.
When Model Outputs Deserve Pushback
Disagreeing with a model is easy; disagreeing with it responsibly is harder. The temptation is to treat your gut as the counterweight, but intuition without a base rate is just noise with confidence. The disciplined move is to ask what the model is actually claiming. If it says a borrower has a 70% chance of default, that number embeds historical patterns that may or may not hold in today's rate environment. Your disagreement should target the reference class, not the conclusion. Are the training cases genuinely comparable to this one? Did the model see conditions like the current regime, or only calm ones? Central bankers' rate predictions fail precisely because the future rarely resembles the averaged past they were trained on.
Once you've located the disagreement, quantify it. Say the base rate is 70% and you believe this case differs — estimate how much, and why. Maybe comparable cases were pre-pandemic, or the lender's incentives shaped the data. A vague "this time feels different" deserves no weight; a specific, falsifiable reason for shifting the prior does. Pushback that survives this test improves the model's next iteration. Pushback that doesn't is just the base rate fallacy wearing a nicer outfit.
Separating Priors From Model Assumptions
The hardest disagreements to articulate are the ones where you accept the base rate but reject the model built on top of it. Suppose historical data shows that 70 percent of startups in a category fail within five years. That base rate is real and should anchor any forecast. But if a particular founder has demonstrably unusual characteristics, you can rationally predict success without denying the 70 percent figure. The trick is being explicit about which part of the reasoning you're contesting: not the prior, but the assumption that this case resembles the reference class.
Most public arguments collapse because participants conflate the two. Say the base rate is wrong and you sound like you're ignoring evidence. Accept it fully and you've conceded the debate before discussing the specifics. The disciplined move is to state your prior, then argue precisely why the case at hand differs from the population that generated it. That shifts the disagreement to something testable — which features matter, how much they shift the odds — rather than a fight over whether statistics apply at all.
Updating Beliefs Without Throwing Out Data
The tension is real: base rates exist because they encode what usually happens, yet sometimes the specific case in front of you genuinely differs. Dismissing the base rate is the classic fallacy; clinging to it blindly is its quieter cousin. The useful move is to treat the base rate as a prior, not a verdict. Ask what reference class the rate comes from, whether your situation truly belongs to that class, and what specific evidence would justify shifting away from it. If you can articulate why your case differs — a mechanism, not just a feeling — you can update, but the burden of proof sits with the deviation.
This discipline matters because ignoring base rates is seductive precisely when stakes are high. Forecasters, borrowers, and policymakers all face the pull of "this time is different" narratives. A practical test: before overriding the base rate, estimate how often people in your position who felt similarly confident turned out to be wrong. If you cannot name the conditions under which you would abandon your dissent, you are not reasoning — you are rationalizing. Disagreement with the model is legitimate only when it is specific, mechanistic, and falsifiable.
A Practical Framework for Model Disagreement
Disagreeing with a model is not the same as dismissing it, yet most people collapse the two into one move. The discipline lies in keeping the base rate intact while arguing about the exceptions. Start by stating what the model predicts and why, then articulate the specific feature of your case that the model's reference class fails to capture. That feature must be concrete enough that a skeptic could check it—not a vague sense that "this time is different," but a named mechanism, a measurable condition, or a documented structural break. If you cannot name it, you do not have a disagreement with the model; you have a discomfort with its conclusion.
Second, quantify the adjustment rather than simply reversing the answer. If the base rate says 70% and your distinguishing feature genuinely applies to a known subset, shift your estimate toward that subset's rate, not all the way to certainty. Third, predetermine what evidence would pull you back toward the base rate. Disagreement without a re-entry condition is just stubbornness wearing analytical clothing. Done well, this turns model skepticism from an act of defiance into an ordinary act of updating—respecting the statistics while taking the particulars seriously.
Ignoring the Base Rate vs. Disagreeing With the Model
| Question | What Ignoring the Base Rate Looks Like | What Disagreeing With the Model Looks Like |
|---|---|---|
| What is the claim? | The base rate is irrelevant because this case is special. | The base rate is real, but the model's update from it is wrong. |
| What evidence is used? | A vivid story, a single anecdote, or a recent headline. | New data, a better-specified model, or a known bias in the original. |
| What happens to the prior? | It is discarded entirely. | It is retained and reweighted. |
| What is the failure mode? | Overconfidence in the exception. | Overfitting to noise or mistaking a model for reality. |