Tails and Uncertainty

What you can and cannot know from data: recognizing fat tails, when averages stop converging, time versus ensemble averages, and two early-warning signals with honest limits.

Six simulations about the limits of inference. The first three show what fat tails do to the ordinary tools — the average, the sample, the expectation. The last three are attempts to detect something before it happens, and each comes with a specific reason the detection is harder than it looks.

Two warnings that apply to this page in particular. The tail estimator generates its own synthetic samples and does not accept your data. The last demo draws a faster-than-exponential curve from parameters you set and does not fit market prices. Neither is a detector, and the second is easy to mistake for one.

Recognizing a fat tail

How would you know a distribution is fat-tailed by looking at it?

On log-log axes, a power law is a straight line. Thin-tailed distributions bend downward because their large events vanish much faster.

Try this. Switch between the two and watch the right-hand end. Straightness on log-log axes is the signature; the downward bend is a thin tail running out of large events.

Assumptions and limits. Straight on log-log is suggestive, not conclusive. Lognormals look straight over a range, and the tail is exactly where you have fewest observations, so the eye is being asked to judge the noisiest part.

Essay: Power Laws, Extremistan, and Non-Ergodicity · Source

When averaging stops working

How many observations does it take before the mean means something?

The Gaussian running mean settles. The Pareto running mean keeps being rewritten by new extremes.

Try this. Run both far enough that the Gaussian mean is visibly flat. The Pareto mean will still be jumping — and each jump is a single new observation rewriting the entire history.

Assumptions and limits. With a tail index below two the variance is infinite and the sample mean has no fixed target to converge to. That is a property of the generator here; establishing it for real data is the hard problem the next demo is about.

Essay: Power Laws, Extremistan, and Non-Ergodicity · Source

Time versus ensemble averages

Is the average outcome the outcome you should expect?

The ensemble mean can rise while the median lived path stagnates. Time averages and ensemble averages are different objects.

Try this. Watch the two lines separate. The ensemble mean climbs on the strength of a few paths, while the path a typical participant actually lives goes nowhere. Both numbers are correct; they answer different questions.

Assumptions and limits. A specific multiplicative process with no ruin barrier. The gap between time and ensemble averages is a property of multiplicative dynamics, not a universal fact about averages.

Essay: Power Laws, Extremistan, and Non-Ergodicity · Source

Estimating the tail index

How precisely can you pin down how fat a tail is?

Move the threshold and watch the Hill tail-index estimate jump. The farther into the tail you go, the fewer observations remain; the estimate gets more relevant and less stable at the same time.

Try this. Move the threshold across its whole range and watch the estimate move with it. There is no threshold at which the answer stops depending on the threshold — that is the finding, not a flaw in the demo.

Assumptions and limits. This generates 500 synthetic Pareto draws with a known tail index; it does not read data you supply. The Hill estimator assumes the tail is exactly Pareto above the cutoff, and trades bias against variance: a low threshold contaminates the estimate with the body, a high one leaves too few points. Here the true answer is known, which is the one luxury real data never gives you.

Essay: The Limits of Knowing · Source

Critical slowing down

Does a system warn you before it tips?

Flatten the valley and the same random knocks produce slower recovery, wider wandering, and higher autocorrelation. That is critical slowing down.

Try this. Flatten the valley in stages. The disturbances are unchanged throughout; what changes is how long recovery takes, how far the state wanders, and how strongly each moment resembles the last.

Assumptions and limits. Critical slowing down is a real signature of approaching a bifurcation, and a weak detector in practice. It also appears without any transition following, it needs long clean records to measure, and a system can tip from a shock large enough not to care how stable the valley was.

Essay: Why the Calm Is Dangerous · Source

Faster-than-exponential shape

What does a bubble look like, and is the shape enough to act on?

Move the growth, feedback, wobble, and critical-time controls. The path bends upward faster than exponential growth, while the oscillations compress as the critical time approaches.

Try this. Raise the feedback exponent until growth outpaces exponential, then watch the oscillations compress toward the critical time. Now move the critical time itself: the curve accommodates almost any value you choose, which is the weakness of the whole approach.

Assumptions and limits. This draws a curve from parameters you set. It does not fit market data. In the words of the source file: the critical-time marker is a fragile model parameter, not a claim that real crashes can be timed from the curve. Fitted on real prices, LPPLS critical times are notoriously unstable across windows — refits move the predicted date, and a moving prediction is not a prediction.

Essay: Faster Than Exponential · Source