Pocket Polymath · Models

Model 001: The Bandwagon Bot

Pocket Polymath now builds its own trading models, and this is the first one. It lives inside the deck as Autopilot, it trades where everyone can watch it, and its record — every win and every loss — is public in the app. We started with the simplest model that could possibly work, on purpose: Model 001 is the baseline every later model has to beat.

What it does

The Bandwagon Bot does what its name admits: it jumps on moves that are already happening. Every ten minutes it studies the day's biggest movers and asks one question about each — is this a wave, or is it chop? The answer is a single number:

z = m / (σ·√k)

where m is how far the price has traveled over the last day, σ is how wildly it wiggled along the way, and k is the number of steps in the window. Trend divided by the noise it had to swim through. A big z is a smooth, steady wave; a small one is a mess that happens to end higher. The bot acts only when conviction clears the bar (|z| ≥ 1.5) — long the clean risers, short the clean fallers.

smooth wave — high zthe bot rides thischoppy spike — low zsame gain, too much noise — skippedBoth end the day up the same amount. Only one is a wave worth riding.

A two-century-old idea

Riding the bandwagon is the oldest documented instinct in speculation. David Ricardo — the economist, trading his own money in London around 1810 — put it as a golden rule: “cut short your losses; let your profits run on.” Charles Dow built the first systematic theory of market trends around 1900. In 1949 Richard Donchian launched the first managed fund built purely on trend-following rules, and in 1983 the famous “Turtle Traders” experiment showed a trend system could be taught to complete beginners — who then made fortunes with it.

Academia caught up late. In 1993 Jegadeesh and Titman published the paper that made momentum respectable — winners keep winning for months, losers keep losing — and it survived three decades of attempts to explain it away. In 2012 Moskowitz, Ooi and Pedersen showed the time-series version (each asset judged against its own past, exactly what our bot does) worked across nearly every market and century of data they could find. Momentum is arguably the most robust anomaly finance has ever documented — and nobody fully agrees on why it works. The leading suspect is human nature itself: people underreact first, then pile on. The bandwagon, in other words.

Crypto, young and sentiment-driven, is where that behavior runs hottest — which is exactly why it's the right proving ground for the naive version of the idea.

Why we ship the naive version first

Every serious modeling effort starts with a baseline: the dumbest honest version of the idea, run in public, so every later refinement has something real to beat. That's Model 001's job. It is deliberately humble — fixed stakes, 1× leverage on real trades, at most three rides at a time, and it only ever risks what you set.

And it must earn trust the only way that counts: in paper mode its every trade is scored on the board, tagged auto, wins and losses alike. If the record is bad, you'll see it — so will we, and Model 002 will know what to fix.

Three temperaments

The bot runs at the aggressiveness you choose. Passive probes: one 15-minute ride at a time, only the very cleanest waves — and after three straight losses it benches itself, because a momentum model on a losing streak is usually telling you the market regime changed. Hold rides up to two waves for an hour each with above-average conviction, and benches itself after four. Bull is the full chase — three rides, your hold setting, no self-stop. The discipline is graded in: the tamer the temperament, the more the model distrusts itself.

Status: the market is winning

Model 001's early live record is losing — which is the system working, not failing: the baseline made its honest attempt in public, and now every future model has a real number to beat. It stays available in the deck with that precaution attached. Model 002 is already in the works, and it changes the question entirely: instead of a formula reading a series, a small vision model will learn from a human's annotated reads of the chart itself — reconciled against what the market actually did.

What could kill it (we're watching)

Educational content, not financial advice. The bot trades only while the tab is open; real-money mode risks real money.

Watch it trade →