The Math Behind the System, and Why It Is Hard to Get Right

I am a fan of computational statistics, which is a polite way of saying I have spent more evenings than I should admit on problems that do not have clean answers. The math inside this system comes from a small set of ideas that are not, on their own, exotic: probabilistic scenario analysis and decision-tree logic, a sequential reasoning architecture and Bayesian-style evidence updating. Anyone with a statistics background recognizes all four. The interesting part is not the names. It is that each one breaks in a specific way the moment you point it at a real stock, and most of the work was learning where.

I am not going to lay out the exact machinery here, and not only for competitive reasons. The honest reason is that the implementation is the least interesting thing about it. What is worth your time is the set of problems it has to survive, because those are the same problems that quietly ruin most analysis, human or machine.

The probability you are quietly inventing

Scenario analysis sounds rigorous. Assign probabilities to outcomes, weight them, take the expected value. The trouble is that the market does not hand you those probabilities. You make them up. Every "30% chance of a re-rating" is a number that walked in from somewhere, and the danger is that the arithmetic which follows looks precise enough to hide how soft its inputs were.

So the first challenge is not computing the expected value. It is disciplining the guess underneath it. A serious process has to force the question of where each probability came from and how much the conclusion moves if it is wrong. Skip that and you have built a very confident machine for laundering a hunch into a forecast.

The branches that multiply

A thesis is never one path. The drug gets approved or it does not, and inside each of those the company raises money or does not, and inside that the macro turns or it does not. Decision-tree logic is the right shape for this, because it refuses to pretend the future is a single line. But trees have a failure mode that is just as real. They branch until they are useless. Model every fork and you drown in scenarios nobody can act on.

The hard part is the pruning. Knowing which forks actually move the outcome and which are noise dressed up as nuance is a judgment the structure can support but cannot make for you. The math tells you how to handle a branch. It does not tell you which branches deserve to exist.

The order that compounds error

Sequence is not a convenience. It is a liability you have to manage. When reasoning runs as a chain, each step inherits the conclusions of the one before it. That is the source of the power, and it is also the trap, because an error introduced early does not stay early. It propagates, and by the final step a small wrong assumption about the business model has quietly bent the valuation, the risk read, and the verdict.

Managing that is most of the difficulty. The goal is to let evidence carry forward without letting mistakes carry forward with it, which means the early steps have to be the most heavily defended, and the chain has to be built so a later step can overturn an earlier one instead of dutifully inheriting it. A sequence that cannot revise itself is just a faster way to be consistently wrong.

The update almost everyone botches

Bayesian-style updating is the part I find most elegant and the part people get most wrong. You hold a prior. New evidence arrives. Your belief should move, and the whole question is by how much. Move too much and you are chasing every headline, repricing the entire thesis on a single data point. Move too little and you are anchored, defending a view the facts have already abandoned. Almost nobody calibrates this well, and that includes models left to their own instincts, which tend to lurch toward whatever they read last.

The challenge is calibration under noise: deciding how strong a piece of evidence really is before letting it move anything at all. Most of what reaches an investor is weak evidence wearing a strong headline. A process that updates honestly has to discount the costume and weigh the thing itself, and getting that weighting consistently right is, in my experience, harder than any single calculation in the system.

Why this is the part worth protecting

None of these four ideas is secret. You could find the textbook treatment of each in an afternoon. What is hard, and what took the time, is applying them honestly to a domain that resists clean numbers, and wiring them together so the weakness of one is checked by another rather than amplified. That is where the real work lives, and it is the part I am content to leave under the hood.

I will say only this. The point of the math was never to make the system look sophisticated. It was to force the analysis to confront the questions it would otherwise skate past: where did that probability come from, which branches actually matter, what does this early step poison if it is wrong, and how much should this new fact really move me. Most analysis never asks them out loud. The math is just what happens when you refuse to stop asking.

Author: Benet Bani