jargAI
Request access
Open menu
Project α · in private research

Everyone's playing. Few see the board.

jargAI sees the board. We don't model what something is worth or how people feel about it. We model the moves people are about to make, learned from the patterns in what they've already done.

Request early accessSee how it's different
  • Not fundamentals
  • Not sentiment
  • Behavior

Our approach

Most models ask the wrong question.

Two people can feel exactly the same and do opposite things. So we skip the opinions and go straight to the behavior.

What we don't

Fundamentals

“What is it worth?”

Slow, public, and already known by everyone who's looking.

What we don't

Sentiment

“How do people feel?”

Cheap to fake and easy to misread. Feelings often never turn into actions.

What we model

Behavior

“What will people do next?”

Habits, timing, and who copies whom. Actions leave patterns, and patterns repeat.

The math

A crowd is a system. Systems can be modeled.

People are more predictable than they think. They show up at the same times, react to the same triggers, and copy the same people. We turn those habits into numbers.

first

Who moves first

One person acts, a few copy, then the crowd piles in. We map that chain and measure how fast it spreads.

spread rate · depth · time to tip

ΔtΔtleaderfollower

Who copies whom

When one account's actions keep showing up in everyone else's a little later, that account is a leader. We measure the delay.

cross-correlation at lag Δt

When they act

Behavior runs on routines: same hours, same days, same triggers. Routines are forecastable.

hour × weekday seasonality

The output is a probability, never an opinion. How likely is this group to act, and how soon? Then we check it against what actually happened.

P(act | history) ∈ [0, 1]

The game

Three kinds of players. Three different moves.

A crowd works like a beauty contest. Winning isn't about being right, it's about knowing what everyone else is about to do. So the question is never just “what are they doing?” but “who is doing it?”

The early

Follow

Acts before it's obvious, and keeps doing it on data they've never seen.

The echo

Ignore

Shows up after the crowd. Loud, sincere, and late.

The ring

Fade

Accounts that keep acting together, on the same thing, at the same time. Their enthusiasm is the tell.

What each move earns you
PlayerFollowIgnoreFade
The earlyWinsNeutralLoses
The echoLosesNeutralNeutral
The ringLosesNeutralWins

Fade means expecting the opposite of what they push.

The catch: every edge fades once it's crowded. When the early get copied, they become the echo, so the model keeps re-learning who is still ahead.

How it works

Four steps, no opinions.

  1. 1

    Watch the actions

    What people do in public: when they show up, what they act on, and who they copy.

  2. 2

    Remove the fakes

    Bots and coordinated rings get flagged and filtered out before anything else.

  3. 3

    Learn the patterns

    Models learn habits, timing, and who leads whom, using only data from the past.

  4. 4

    Predict the next move

    A probability that a group acts, and when, checked against what actually happens.

project-α

$ jarg watch --public

▸ actions loggedlive

▸ fakes filteredon

$ jarg learn --habits --lead-lag

▸ leaders mappedrunning

$ jarg predict --next-move

▸ outputP(act) in [0, 1]

▸ beats baselinetesting

$ ▍

Rigor

We try to break it before you do.

Every prediction is scored by how well it ranks what really happened, on data it never trained on. A score of 0.50 is a coin flip. A number only matters if it survives these six rules.

  1. 01

    Test on the future

    Models learn from the past and get tested on a later stretch they've never seen, over and over.

  2. 02

    A locked vault

    Everything after a fixed date stays out of training. It's only used to check whether a result holds up.

  3. 03

    No peeking

    A signal only counts if it existed before the moment of the prediction. Anything recorded later gets thrown out.

  4. 04

    Beat the obvious

    Every data source has to beat a simple baseline. If it doesn't add anything on new data, it doesn't ship.

  5. 05

    Honest error bars

    Confidence ranges resample whole days, because everything on the same day tends to move together.

  6. 06

    Publish the misses

    Most patterns that look real turn out to be noise or something obvious. When that happens, we say so.

Want in before it's obvious?

We're opening Project α to a small group of early partners. Leave your email and we'll reach out when we can share more.

Limited spots. No spam.