Most traders treat their conviction like a light switch: fully on, or fully off. A position either "works" or it "doesn't work." A thesis is either right or wrong. This binary reflex produces a particular kind of error — holding a crumbling view far too long, then abandoning it entirely just before it stabilizes. The skill this article is about is the alternative: treating conviction as a dial, and learning to turn it the right amount each time new evidence arrives.

By the end of this article, you will be able to apply a three-step update process to any thesis, distinguish evidence that genuinely shifts the probability from evidence that merely confirms what you already believed, and recognize the all-or-nothing updating pattern before it costs you.

Beliefs Are Dials, Not Switches

A probability is not a prediction. It is a statement about how much weight a specific hypothesis deserves, given everything you currently know. When you enter a position with "65% conviction," you are not saying it will work — you are saying that, on balance, the evidence available to you right now supports the thesis more than it undermines it, and 65 is your honest calibration of that balance.

The dial metaphor matters because it forces two uncomfortable disciplines. First, you must actually put a number on it — vague confidence does not survive contact with contradictory data. Second, you must move the dial in proportion to the new evidence, not in proportion to how the new evidence made you feel. These are different things, and the gap between them is where most updating errors live.

The Three-Step Update

Bayesian updating in a trading context reduces to three questions, asked in sequence every time new information arrives.

Step 1 — State your prior

Before you evaluate the new piece of information, write down your current conviction as a percentage. "I believe this thesis is intact at 70%." That number is your prior. If you skip this step, the new information will anchor you without you noticing. You will feel as though you weighed it objectively, but you will have processed it against an invisible, unstated starting point that may already be distorted by recent losses or recent wins.

Step 2 — Ask the diagnostic question

The diagnostic question is: How much more likely is this piece of evidence if my thesis is true versus if it is false? This is the question most traders never ask, and it is the one that matters most. If the evidence is equally likely to appear whether your thesis is correct or not, it carries no information. It does not belong on your dial. If the evidence is significantly more likely under one scenario than the other, it is diagnostic — it should move the dial.

Step 3 — Move the dial in proportion

Weak evidence moves the dial a few points. Strong, discriminating evidence moves it significantly. The move should be proportional to the answer you gave in step two. A piece of data that was only slightly more consistent with your thesis than with the alternative adds perhaps 3–5 points of conviction. A piece of data that is dramatically more consistent with your thesis — that you would not expect to see at all if you were wrong — might justify 15–20 points of movement.

The updating rule is not mechanical; it requires honest judgment. What it eliminates is the binary impulse: the new data point that makes you abandon a 70% conviction overnight, or the irrelevant noise that inflates a 40% hunch to 90% without earning it.

What "Diagnostic" Actually Means

Understanding diagnosticity is the hardest part of this skill, and also the most neglected.

Consider three types of evidence a trader might encounter during a thesis:

Non-discriminating confirmation. The asset you hold is down on a day when the broader market is also down. Does this confirm your short thesis? Not meaningfully — because the same thing would happen whether your specific thesis is right or wrong. The market selloff is not a diagnostic test of your particular view. Treating it as one inflates your conviction with noise.

Weak evidence. The company you hold reports revenue slightly below analyst consensus. This is mildly more consistent with a weakening thesis than with a strengthening one, but the miss is small and within historical variance. A few points of dial movement may be justified. More than that is disproportionate.

Strong evidence. The same company announces that its largest customer — the one your thesis was built around retaining — has terminated its contract. This is evidence that is dramatically less consistent with the bull thesis than with the bear. The dial deserves a large move, quickly. Holding 65% conviction through that announcement while telling yourself to "stay patient" is not discipline — it is a refusal to update.

The practical rule: before you let a piece of information move your dial, ask whether it would look different depending on which scenario is true. If the answer is "not really," file it, do not use it.

Historical Example: LTCM, 1998

Long-Term Capital Management came to embody a different kind of updating failure — not stubbornness, but a structural inability to generate an honest prior.

LTCM's models were built on correlation assumptions derived from years of historical data. Those assumptions implied that certain simultaneous, large moves across global markets were extremely improbable. One reading of the episode is that the models did not treat this as a tentative prior subject to revision — they built position sizes around it as if the probability were essentially fixed. One way to read the episode is that the models had no real mechanism to update; they implied near-zero probability for events that were merely rare under normal conditions, and rare-under-normal is not the same as rare-under-stress.

On August 17, 1998, Russia announced a moratorium on its domestic debt and devalued the ruble. What followed was a rapid, correlated repricing across global credit and equity markets. As Alan Greenspan later described in congressional testimony, the crisis environment was "so at variance with the experience built into its models" that the correlation assumptions that underpinned LTCM's positions collapsed together. Within weeks, the fund faced losses across multiple positions simultaneously — not because each trade was wrong on its own terms, but because the correlations the models assumed would stay low spiked toward one.

In September 1998, a consortium of 14 private banks completed a recapitalization of approximately $3.6 billion. The New York Fed organized the process but provided no public funds — Greenspan testified explicitly that it "was not a government bailout." It was a private-sector recapitalization coordinated by the New York Fed.

The updating lesson is narrow but precise: a model that assigns near-zero probability to an event and then encounters that event cannot update gracefully, because it has left itself almost no room. A dial that starts at 1% and should move to 60% has no vocabulary for the distance. Keeping a prior further from the extremes — acknowledging that rare-but-not-impossible events exist — is not pessimism. It is what preserves the ability to respond proportionally when they arrive.

The Failure Mode: All-or-Nothing Updating

All-or-nothing updating is the most common version of this error. It operates in two directions.

The freeze. A trader holds conviction at 80% through accumulating counter-evidence, treating each piece as insufficient to justify a change. Each individual piece is indeed insufficient — but the cumulative drift of five moderate signals is not. The dial should have moved gradually from 80% to 50% to 35% over several sessions. Instead it stays at 80% until the evidence is so overwhelming that the trader capitulates entirely — now at 10%, not 40%. They moved from overconfident to pessimistic without pausing at calibrated.

The flip. A trader abandons a thesis the moment a single piece of bad news arrives, even if that news is weak or expected. The dial drops from 70% to 5% overnight. This feels like discipline — "I took the loss." But it is the same failure mirrored: the update was driven by emotional response to loss, not by the diagnostic weight of the evidence. If the evidence warranted a move to 50%, the move to 5% means the next version of this position, when conditions normalize, will be missed.

The emotional tell for both versions is the same: the conviction moved proportionally to how the trader felt about the evidence, not proportionally to what the evidence actually discriminated. Scenario pre-writing helps catch this — the article Scenario Calibration and Triggers covers how to define update triggers before you need them, which is a complementary discipline to what is described here.

A related trap worth naming: confirmation bias. Confirmation bias is what happens when you selectively gather evidence for the scenario you want to be true. It is a filter on what enters your updating process, whereas all-or-nothing updating is an error in how you process what does enter. Both impair calibration, but they operate at different points. The article Confirmation Bias at the Chart covers the filter problem directly.

The Discipline: A Pre-Session Update Log

Keep a one-paragraph updating log for any thesis you are holding. Before each session, record your prior. During the session, note the two or three most significant pieces of new information you encountered. For each, answer the diagnostic question — would I expect to see this under the alternative scenario? Then record your new conviction and a one-sentence explanation of why you moved it the amount you did.

The log does not need to be long. It needs to be honest. A record that reads "moved from 65% to 62% because the data was weak" is more valuable than a record that reads "staying patient" four times in a row before a capitulation to 10%. The trail of small, proportional moves is evidence of a developing skill. The trail of flatlines and jumps is a pattern worth confronting.

Simulator Exercise: The Conviction Dial

Open Abu Terminal and start a Speed Run in any era. Before the first event plays, set a written conviction level for the underlying thesis in that scenario — assign a percentage between 10 and 90, and write it down or say it aloud.

As events surface sequentially, apply the three-step process to each one. Ask: is this evidence diagnostic? If yes, move your conviction by the amount the evidence warrants — not more. After each event, record your new conviction and a one-word description of why it moved (or why it did not).

The drill specifically rewards proportional updating. A jump from 60% to 95% on a single weak signal, or a drop from 60% to 5% on a single moderate one, is the behavioral fingerprint the drill is designed to surface. The simulator is not judging your final score — it is giving you a safe environment to notice when your updates are being driven by the narrative energy of a sequence rather than by the actual diagnostic weight of each event. After the run, review your update log: were the large moves earned by strong, discriminating evidence, or by emotional momentum?

If you find that your conviction tends to march steadily in one direction regardless of the evidence quality of individual events, that pattern is worth sitting with. It suggests the dial is tracking the recent trend of outcomes rather than the strength of each piece of information — which is a different skill problem, covered in Base Rates and Priors.

Bayesian Updating: How to Change Your Mind Without Whiplash