Signals

Not investment advice. This is the output of a personal model made by a private individual. It is not a recommendation to anyone to buy or sell any security. The author may hold positions in some of the securities mentioned.

Signals trial & error dataset

Operation: Millionaire. Trial. Profit. Error.

I need money. My goal is €1,000,000. My plan: test money ideas in public, with real data, and laugh at myself when they fail.

The bill so far

  • Server €35/month
  • Claude Max $100/month
  • Revenue €0
The full bill ↓

Week 1. Ideas tested: 0. Confidence: suspiciously high.

See the plan → ☕ Buy the nerd a coffee

Tips are voluntary and buy nothing — no picks, no advice, no access. Just caffeine and server time.

Cartoon of the Number Nerd, arms crossed, smirking in front of a wall of dashboards and a query that returns 0 rows A desk monitor running a data notebook: a scatter plot with a limp regression line, a histogram, candlesticks and a terminal. Ones and zeros rain in the background, a database sits beside it, and a chart line escapes the screen, climbs, then crashes. 1011001 0110100 1001011 0101101 1100010 alpha_hunt.ipynb > model.fit() -- 0 rows IDEA #1SOON GRAVEYARDEMPTY TO DOGET RICH

The roadmap (goals only)

The road to €1,000,000

Here's the climb. Spoiler: I'm at the very bottom, at the data bench, with a coffee and an empty stack of autopsy reports.

  1. You are hereNowWeek 1

    Needing money

    0 revealed, 0 buried, 5 coming. Coffee: in progress.

  2. GoalCheckpoint 1

    A model that beats a coin flip

    On its own test, with the rules written down first. A low bar. Nobody has tried to clear it yet.

  3. GoalCheckpoint 2

    12 months on the record

    Model 05, one recorded month at a time. Patience is the whole strategy.

  4. Goal€100,000

    Six figures

    A checkpoint, not a forecast. Very much not a forecast.

  5. Cartoon of the Number Nerd fist-pumping and winking at a laptop that reads p < 0.05The goal€1,000,000

    Millionaire

    Personal goal. Keynes supposedly warned that markets can stay irrational for longer than you stay solvent. So: wish me luck.

Every checkpoint on this road is a goal, not progress. I don't publish a portfolio size, and none of this is a forecast or a promise, to you or, frankly, to me.

The dataset

Big data in. Ideas: pending.

I don't have money, but I do have a pile of data, counted straight from my own server. Brace yourself.

Just counting: Read one per second and you need about 9 months. No sleep.

23,318,052 daily stock prices

Just counting: Just reading their names out, one per second, takes nearly 3 hours.

10,433 stocks

Just counting: Longer than a 13-year-old has been alive.

16 years of daily prices, Apr 2010 – Aug 2026

Just counting: Short selling is betting a price will fall. Eight years of other people's gloom.

10.7M rows of short-selling volume, Aug 2018 – Aug 2026

Just counting: One per second, non-stop: about 12 days.

1,066,026 facts from company accounts, 3,966 companies

Just counting: It starts before most grandparents were born.

100 years of monthly market data, 1926–2026

23 million prices, 10,433 stocks, 16 years. Ideas tested so far: zero. The pile is huge. The plan is a sticky note.

Show me the data

> SELECT table, rows FROM engine ORDER BY rows DESC;

Rows per table, log scale Log-scale bar chart of the row counts in the table below, from 23,318,052 daily prices down to 1,200 monthly factor rows. Each grid line is ten times the one before. 1001k10k100k1M10M100M daily_prices short_volume short_interest xbrl_facts fred_series factor_rows
tablerowsspan
daily prices23,318,05210,433 tickers, Apr 2010 – Aug 2026
FINRA short volume10,721,689Aug 2018 – Aug 2026
short interest3,866,270Jan 2018 – Jul 2026
XBRL fundamentals facts1,066,0263,966 companies
FRED economic series rows1,648public economic data
monthly factor rows1,2001926 – 2026, Ken French data library

For nerds: counted read-only in October 2026, data up to 8 Aug 2026. XBRL = the machine-readable version of company accounts. A factor = a trait like cheapness or recent winning that people rank stocks on. Everything on the server disk: about 35 GB, backups included, because paranoia is also a dataset.

full source list → Method

The trial & error board

Five models coming. Zero revealed.

Week 1 is for admitting I need money. The ideas come next, one at a time, and each one that turns into a model goes public on its own reveal. Plain answer on top. The real data under “Show me the data”. No rounding up, no cherry-picking, no “it would have worked if”. For now, every card is a closed envelope.

Model 01
Reveal: soon

Coming

The first idea

The teaserFirst idea: arriving soon. Probably bad.

Answer
? Nothing to answer yet. First the idea, then the rules, then the test.
The teaser →

Model 02
Reveal: after 01

Coming

The sequel

The teaserA smarter way to read the past. Or at least a longer one. Does it find anything?

Answer
? Sealed until Model 01 has had its turn. No spoilers, not even a wink.
The teaser →

Model 03
Reveal: after 02

Coming

The rebuild

The teaserRebuilt from the ground up and run again. Does a clean rebuild change the answer?

Answer
? Sealed until Model 02 has had its turn. The envelope stays shut.
The teaser →

Model 04
Reveal: after 03

Coming

The grown-up test

The teaserWhat happens when you show a model to a proper factor test?

Answer
? Sealed until Model 03 has had its turn. Patience is also a factor.
The teaser →

Model 05
Reveal: at launch

At launch →

The Provable Allocator

The teaserA slow, boring, monthly model with the rules written down first. How boring, and whether boring works?

Answer
? Not known yet. The details go public at launch.
The teaser → Coming soon

The feed

Latest from the feed

All posts →

More info

The small print, for the curious.

The rules, the graveyard, the commit log, the bill and the archive. All still here. All closed until you open them.

  • The graveyard of hypothesesNobody has died yet. Five plots are reserved.
    Cartoon of the Number Nerd sipping coffee, deadpan, while his dashboard goes red, the kernel dies and a FAILED stamp lands

    Nobody has died yet. Give it a week. When a model fails, it gets buried here with full honours, and its pages stay in the archive, untouched: quietly deleting the losers is how survivorship bias starts. Five plots are reserved. The groundskeeper calls that planning ahead.

    • Model 01

      Status: idea pending

      Plot reserved. Shovel ready, just in case.

    • Model 02

      Status: in the queue

      Plot reserved. Booked, not yet occupied.

    • Model 03

      Status: in the queue

      Plot reserved. Name on the list, not on the stone.

    • Model 04

      Status: in the queue

      Plot reserved. No comment until the reveal.

    • Model 05
      Provable Allocator

      Status: at launch

      Plot reserved, paid in advance. Hopefully not needed.

  • Week 1: one goal, zero models, no champagneThe commit log so far.
    Cartoon of the Number Nerd hammering a laptop with an unhinged grin while code and data stream out of it

    1. Operation: Millionaire starts

      I need money. Goal: €1,000,000, my own and nobody else's. Plan: test money ideas in public, with real data, and laugh at myself when they fail. Ideas tested so far: 0.

    2. This notebook goes public

      A scoreboard, a bill and a graveyard with no residents yet. The content calendar is the only part of this operation that runs on time.

    3. Next

      The first idea

      Coming. Probably bad. Definitely tested.

    4. After that

      Model 01 goes public

      The bet, the rules, the test, the verdict. In that order.

    5. After that

      Models 02 to 04

      One at a time. No skipping the queue.

    6. At launch

      Model 05 on a monthly track

      The Provable Allocator, one recorded month at a time. Slow, boring and very hard to fudge.

  • What the tuition boughtThree rules, learned in advance. Cheaper that way.
    Cartoon of the Number Nerd, fist under his chin, squinting skeptically at a scatter plot with a limp regression line

    No tuition paid yet. The market bills in time and ego, never in advance. These are the rules I promise to remember before it teaches them to me the expensive way; real lessons get added as the models go public.

    • Write the rules down before you look.

      Settings fixed in advance, run once, no tuning on the result. That's the only way a "no" means something, and the only way a future "yes" will.

      Source: Method

    • Only test on news you could have read.

      Anything faster is look-ahead bias: a backtest that reads tomorrow's paper. It always looks brilliant, and it is always fiction.

      Source: Method

    • Show the bad results too.

      Especially the bad results. A notebook that only shows winners is an advert.

      Source: Method
  • P&L so farWhat this costs to run, and what it brings in.

    A notebook that publishes its failures should publish its running costs too. Here's what Operation: Millionaire costs to keep alive, and what it brings in.

    Revenue
    €0 So far. The tip jar is open now; this line is still very easy to audit.
    Server
    €35/month for the server that runs the models, with its data volume.
    Claude Max
    $100/month (Max 5x, billed in dollars) for the AI that writes my code and my excuses.
    Monthly burn
    ≈ €35 + $100 a month. Two currencies, no conversion. I'd rather not do FX maths on a loss.

    Net result: a monthly loss, paid in two currencies. Very on-brand.

  • The archive: four shelves, all under sealThe evidence locker. Nothing comes out before its reveal.

    Every model's raw output will be kept here exactly as it was generated. No retouching, no hindsight edits. Right now every shelf is sealed: nothing comes out until its model goes public. Times in UTC. For the plain-English version of each model, start with the model cards.

    • Vol1Under seal until Model 01 goes publicComing · soonNo spoilers in the evidence locker either. Back soon.
    • Vol2Under seal until Model 02 goes publicComing · after Vol 1No spoilers in the evidence locker either. Back soon.
    • Vol3Under seal until Model 03 goes publicComing · after Vol 2No spoilers in the evidence locker either. Back soon.
    • Vol4Under seal until Model 04 goes publicComing · after Vol 3No spoilers in the evidence locker either. Back soon.

    Heads up: the shelves are empty. Failure, it turns out, takes a while to document.

Tip jar

Subsidise the science

The models run on a server, an AI subscription and stubbornness. If this notebook made you laugh, or talked you out of a bad idea, you can help cover the bill.

  • Coffee for late-night backtests
  • Helps pay the server and the AI on the bill above
  • Changes nothing about what gets published. Failures included. Especially failures.
☕ Help cover the bill

Tips are voluntary and buy nothing — no picks, no advice, no access. Just caffeine and server time.

A steaming coffee cup labelled data fuel, with a small rocket taking off next to it DATA FUEL

Follow along

Follow the journey. One idea at a time.

The first idea goes public here soon, then the models one by one, and Model 05 at launch. Back soon. No newsletter, no sign-up, no forms, no “exclusive community”. Just come back and check the scoreboard. Popcorn optional.

Back to the models Read the method

The €1,000,000 is my personal goal. This site is a record of my own experiments. It is not advice, and nothing here is a recommendation to buy or sell anything.