This article talks about ways to deal with uncertainty, and I will also briefly define the concept in contrast with simple risk.

- Definitions: Uncertainty vs Risk
- Rational Estimation without Uncertainty
- Rational Estimation with Uncertainty

**1 – Definitions: Uncertainty vs Risk**

- Uncertainty and risk are both not simple economic phenomena. They are generic elements of complex systems.
- They are both represented as error in a forward-tested model, but only risk is captured in a backward-tested model.
- Risk
- A probability of some event
- Sometimes, the expected value of some event: The probability multiplied by the payoff
- Think of your budget: You have a line item for gas, but there is some risk you will go a bit above or below that amount.
- You are familiar with the kind of event
- You may have bought insurance for this thing because you knew beforehand it might happen

- Uncertainty
- Synonyms
- Think of your budget: You have a category for “Other” or “Unexpected Expenditure”
- You probably didn’t buy insurance for this thing because you don’t know what it is and no one sells “Other Stuff Insurance” (yet…?)

**2 – Rational Estimation without Uncertainty**

Before we go on to estimating the costs of uncertainty (which I argue can be done), what are some principles of estimation when uncertainty is not the case?

- If you have lots of data, do a regression. If you don’t, take an average. You never have no data.
- The regression should be linear by default unless you have some reason for another approach.

- Direction-ambiguous bias doesn’t invalidate or prevent estimation; it preserves the point estimate.
- If I could spend more, less, or the same on gas as I budgeted, should I throw out the budget? No! It means I budgeted the optimal estimate.
- Obviously, if you have some reason to think the weight one way is bigger then you can apply some correction. But if you just don’t know the errors offset in either direction.
- More error in both directions will reduce our confidence in our point-estimate, but it doesn’t shift the value of the point estimate.

- Direction-specified bias can be corrected using ordinal logic
- If I know I will spend more on gas next month, how should I proceed?
- Ask yourself, “How much more?” Based on your answer:
- “I don’t know.” Assume no change. This is also an axiomatic way to justify the default usage of the linear regression.
- “A small amount more.” Small relative to what? The original value. New budgeted value = 1.5(Original)
- “An even amount.” New budgeted value = Original
- “A large amount more.” New budgeted value = 2(Original)

- I have a long personal literature on ordinal modelling for more detail including the maths:

- Ask yourself, “How much more?” Based on your answer:

- If I know I will spend more on gas next month, how should I proceed?
- Abstraction reduces the accuracy of estimation, but doesn’t prevent the possibility of estimation
- Corollary: The accuracy of a forecast is roughly commensurate with the number of observable input or input properties
- This is why micro-founded economic models are more accurate than aggregate models. Eg, including individual level characteristics improve accuracy
- You can do a bit of handling for unobserved heterogeneity, but not much. You pretty much get 1 extra parameter in addition to observable data.

- Any set of observations, events, or items can be abstracted to some level such that they belong to a common class.
- Any item belonging to a common class can be estimated on the grounds of the other items belonging to the class.
- So a maximum-likelihood estimate of any item can be obtained. The tricky part is reducing the error such that substantial confidence or usefulness is gained.

- Corollary: The accuracy of a forecast is roughly commensurate with the number of observable input or input properties
- Information contained in the model is represented in the portion of variation explained.
- Corollary: Ignorance is represented as a smaller share of variation explained, more error in forecasting, and/or less confidence in estimation of various parameters.

**3 – Rational Estimation with Uncertainty**

I argue that case probability estimation is possible, contrary to what Mises has said, for two reasons:

- Everything can be classified at some level of abstraction, even if it is a retrospective category for “Unforeseen costs” or “Other”
- If you have never suffered an unforeseen cost then you:
- Probably shouldn’t budget for it, or
- Budget for it at the rate of a foreseen cost since this is your only rational focal point / Schelling point

- If you have some data, even if it is small-n but definitely for large-n, on the frequency and payoff of unexpected events then use that.

- If you have never suffered an unforeseen cost then you:
- It is possible to demand surprise
- Or, exert effort or money to influence the own-rate of surprise or the rate of discovery
- There are things I know alot about and things I don’t know alot about
- If I engage in an activity I know very little about it is easy for me to be surprised
- Eg visiting China: Known unknown = I have to learn Mandarin.
- Unknown unknown = I don’t know!
- But I can set aside some cash for it just in case
- More cash set aside reduces the potential negative consequence
- At the cost of foregoing known rewards

- If I engage in an activity I know alot about it is hard for me to be surprised
- Eg talking about economics: I have probably already heard it and even if it’s new I doubt it’s highly surprising
- Or, getting a desk job and never doing new things. The same tasks over and over. No learning, no discovery.

- If I engage in an activity I know very little about it is easy for me to be surprised
- A friend suggested that things I know alot about actually increase my ability to be surprised: Fine, it’s possible, but I was still able to influence the own-rate of surprise

- Tangentially, it should be possible to estimate the rate of innovation whether such rate is endogenous or totally random.

Here are some applications:

- Budget for unknown expenditures
- Include an “Other” category in surveys
- In Agile planning, you can plan using points which are associated with unknown requirements
- Give Rumsfeld extra money even though he doesn’t have a specific thing to spend it on (I’m not really recommending this, but it’s the conclusion of his remark)
- When investing, double the risk premium to include a rational estimate of Knightian Uncertainty
- Note: Don’t assume an increase to risk unless you have a reason. Knightian Uncertainty could reduce you risk.
- Another Note: Rational doesn’t actually mean true or accurate. It just means you got nothing better. It’s definitely better than the idea that estimation is impossible.
- A Third Note: But like, what are you going to do, allow for infinite uncertainty? That’s not rational or solvable, but we assume uncertainty all the time. Uncertainty is solvable.

- Key point: It is possible to estimate the frequency and cost (or benefit) of a surprise, but it is not possible to precisely identify the kind or content of a surprise.
- In that case it wouldn’t be a surprise; the precise content is revealed in the moment of discovery.
- But, it may be possible to estimate the broad nature of a surprise.
- Eg there is an emerging field of medicine likely to bring more inventions compared to legacy fields.
- Very hard to do at a social level as it stands; maybe not too hard for a self-reflecting individual to recognize high potential areas of personal discovery or surprise.

## 2 thoughts on “Rational Estimation and Price Under Uncertainty”