prior/odds

Prior odds. Posterior work.

Let the evidence decide.

I'm Elvis Liu — recovering internal auditor, now risk and forecasting. I build small self-hosted systems and treat prediction markets as the lens for everything.

Flip the coin Read the books

The chart

The idea, in one coin.

Start from a wide guess, let each flip move it, then reveal the truth and see if your range caught it.

A coin with a hidden bias
The chance of heads runs left to right, from 0 to 1. A dashed curve is the prior: a wide guess centred on 0.5. A solid curve is the current belief; it narrows as coin flips arrive. belief 0.50 prior
Flips
0
Heads
0
Belief
0.50
90% range
0.14–0.86

Before any evidence: a wide prior, centred on a fair coin.

Same evidence. Different minds.

The flips above moved one belief. Now start from a hunch of 30%, decide how sure you are, and let the same twenty flips land.

00.250.50.751

Hunch 0.30, worth 10 flips. After 20 real flips (14 heads): 0.57 — moved +0.27.

Middle ground: the evidence moves you, but the hunch still counts.

Get the highlights.

Calibration over confidence.

Every call logged, scored, and left to age in public. On the diagonal, your 70%s come true 70% of the time.

Post coming soon

Three books. One habit.

Data, AI safety, and accounting — each taught through one story, every number checked.

Learn more

Self-host the loop.

A private notebook, a forecasting brain, a chat gateway. No SaaS bill, no tracking.

Post coming soon

Read odds like a forecaster.

Prediction markets as a lens: the world as live probabilities. Risk control, but for everything.

Post coming soon

Agents that ship.

Claude Code as a teammate — tools, loops and harnesses that do real work.

Post coming soon

Not louder. Better calibrated.

How I think.

  1. Calibration over confidence.

    A loud take is worthless if your 70%s only land half the time.

  2. Small bets, all logged.

    Every call gets a number, a date, and nowhere to hide.

  3. Self-host the loop.

    Own the tools and the data. No SaaS between me and the work.

  4. Evidence over vibes.

    Update when the world pushes back, not when it feels good.

  5. Wrong in public.

    A scoreboard only counts if you can't quietly edit it later.

Three books.
Plain words, two languages.

Cover of Steeped in Data: two teacups beneath a teal and an orange bell curve English · 中文 Steeped in Data Data analysis, pipelines and A/B testing — one case to solve, with live SQL in your browser. Find the missing 12%, and learn the craft on the way. Read now Cover of Inside the Guardrails: a closed book in a circle of lamplight, at night English · 中文 Inside the Guardrails How today's AI models are made safe — one fictional lab, a small real model, every number measured. Keep one secret for six weeks, and learn how AI safety works. Read now 中文 · Parts I–IV One Ledger China's six CPA exam subjects and financial analysis as one subject, taught through one fictional company. The prologue and Parts I–IV (to Ch 101) are online, updated for the 2027 exams. Read now
  • 3books, written in plain language
  • 152chapters online — 24 + 27 + 101
  • 2languages: English and 中文
  • 0trackers, cookies or paywalls

What I run.

All of it self-hosted. No SaaS between me and the work.

  • Notebook Where the thinking lives. Plain markdown, versioned, mine.
  • Brain Forecasting brain. A retrieval layer over everything I've read and saved.
  • Gateway Chat gateway. One self-hosted door to the models I actually use.
  • Workspace Code workspace. A browser IDE, so I can build from anywhere.
  • Site This site. Astro, hand-built, SVG. No tracker, no cookie.

Wrong in public.

Featured priors, prior → posterior over seven days. The gap is the lesson, and it's left to age.

  • PM:RATE-Q3 Fed cut by Q3 0.42→0.55 +0.13
  • PM:CLAUDE-5 Anthropic ships Claude 5 EOY 0.62→0.71 +0.09
  • PM:AGI-EOY Frontier lab claims AGI EOY 0.18→0.12 −0.06
  • PM:ELEC-26 US 2026 incumbent party retains 0.54→0.61 +0.07

Keep reading.

All entries

Bet with me,
not against the evidence.

Working on forecasting, risk, or self-hosted AI? I like a good prior and a sharper posterior.

hi@priorodds.com Chat GitHub