The method, short version (for now)
How I'll test ideas
No model is public yet, so this page is short. These are the rules, written down before anything gets tested. Each model adds its own method notes here when it goes public. The jokes live on the homepage; this is the notebook.
Signals (Docka Signals) is a public notebook of financial model experiments by the Number Nerd. Not investment advice.
First public release: 2026-10-16. Rules written before results. The owner, the court and the defendant are the same person. Separation of powers is on the roadmap.
Plain facts
What this site is
- Every idea becomes one sentence that could be wrong, and its test rules are written down before it is run.
- Every result is published and scored against a coin flip, which gets 50% on “up or down?”.
- Failed models stay in the archive exactly as they were, with no edits after the fact.
In one paragraph. I need money, and my personal goal is €1,000,000 (a goal, not a forecast or a promise). The plan: take one money idea at a time, turn it into a sentence that could be wrong, write the test rules down first, run the test on real data, and publish the result, good or bad. No model is public yet. The first idea arrives soon. Shorter still: zero models, five coming, one very patient coin.
Written down first
The rules, before any test
- Write the bet down. One sentence that could be wrong. If it can't be wrong, it isn't a model. It's a horoscope.
- Fix the rules first. The scoring rules and settings are set before the test and not changed after a peek. Changing them after a peek is called p-hacking: torture the data long enough and it confesses to anything.
- Only use what I could have known. A test may only use news that was public at the time, and acts at the next market open after it, never the same day. Using news before you could have read it is called look-ahead bias.
- Score it against a coin. Every result is shown next to the base rate: what pure guessing would score. On "up or down?" a coin gets 50%. A model that can't do better than a coin doesn't get to brag.
- Keep it or bury it. No edge, no mercy. A failed model goes to the graveyard and its pages go to the archive, exactly as they were. No edits after the fact.
No hiding
Bad results get shown too
Every result goes on the page, including the bad ones. Especially the bad ones. Quietly deleting the losers is how survivorship bias starts: a record that forgot its failures, so every result looks better than it was. Here the failures stay, with their dates. I don't publish returns, profits or a portfolio size, only whether a model did better than a coin on its own test.
The pile
The data on hand
Counted read-only from my server's databases in October 2026, data up to 8 Aug 2026. Which data each model uses is written up when that model goes public.
| Source | What | Rows | Years |
|---|---|---|---|
| Public market data | Daily prices, 10,433 tickers | 23,318,052 | Apr 2010 – Aug 2026 |
| FINRA | Daily short volume | 10,721,689 | Aug 2018 – Aug 2026 |
| Public short-interest data | Short interest | 3,866,270 | Jan 2018 – Jul 2026 |
| SEC EDGAR (XBRL) | Company fundamentals facts, 3,966 companies | 1,066,026 | as filed |
| FRED | Economic data series | 1,648 | — |
| Ken French data library | Monthly factor returns | 1,200 | 1926 – 2026 |
Everything on the server disk: about 35 GB, backups included. Big numbers in; the results come later, and they will be shown whatever they are.
Not advice
This page describes how I test my own ideas. It is not advice on what you should do with anything, and nothing here is a recommendation to buy or sell any security. The €1,000,000 is my personal goal, not a forecast, not a promise, and not a target for anyone else. The full legal text is on the disclaimer page.