EXP-014 logged Apr 9, 2026
Small-cap overnight gap fade
The bar: average 5-day return above 0.4%, judged on data the model never saw.
Verdict: FAILED. The effect disappeared on fresh data.
On the record
Entry 001
Opened Jul 13, 2026
V
Teaching a neural network to trade, in public.Every experiment published, win or lose.
A one-person lab building custom neural networks to predict short-term stock moves. Every experiment is written down before it runs; every verdict is posted unedited, losses included.
what runs here
- A custom neural network ensemble, built by one trader, watches the market minute by minute, every session.
- When it predicts a +5% move, the signal fires in real time.
- A machine grades every flagged signal overnight.
- Follow along at The Graded Tape by morning, wins and misses alike.
6,141 signals graded
0 withheld
graded by code each morning
00
The system
Current state
V
The loop runs every night: the models retrain on yesterday’s data, and the tape re-posts by morning.
Training runs · six members walking forward
v1_championv4_twoheadv6_prunedv8_noewcsig_m32_R512logistic
01
The record
Exhibit A
V
The Graded Tape
A machine-kept record, hosted at gradedtape.com. Every morning, code scores yesterday’s signals against rules fixed in advance and writes them to a tape it can’t edit after the fact. Signals are grouped into bands by share price. Three of four price bands sit below zero, and they stay on the page.
02
The problem
Exhibit B
V
Most published track records are filtered before they reach you.
The winners are listed, the losers are gone, and the strategy quietly changed the month it stopped working.
Exhibit B is a track record assembled the usual way. The shaded months were left out of the brochure.
01+3.1%
02−4.8%Left out
03+1.9%
04+2.6%
05−6.2%Left out
06+4.4%
07+0.8%
08−3.7%Left out
09+2.2%
10−5.1%Left out
11+3.5%
12+1.4%
The brochure shows 8 of 12 months: +19.9%
All 12 months together: +0.1%
03
The build
Committed before results
Illustrative · format examples
V
Written down first. Run second. Published either way.
- 1 · The experiment is written down and timestamped before it runs; the write-ups are published as letters.
- 2 · It runs.
- 3 · The result is published unedited, win or lose.
Each experiment’s write-up names which stocks, how long the trade holds, what counts as a win, and how it’s scored. The experiment runs once, on data the model has never seen. The verdict is posted the same way every time, no matter what it says.
EXP-015 logged May 2, 2026
Post-earnings drift, liquid names only
The bar: the highest-confidence signals hit their target more often than a preset bar.
Verdict: PASSED. It cleared the bar on fresh data.
EXP-016 logged Jul 1, 2026
3-day momentum versus peers
The bar: written and sealed Jul 1, 2026.
Verdict: OPEN. Grading from Aug 15, 2026.
These cards show the format: every experiment written down and timestamped before it runs. The graded results are on The Graded Tape. A win only counts against the full count of tries.
04
Method
Fixed in advance
V
Why this record can be trusted.
01
Judged only on data the model never saw
Every claim rests on data the model never trained on. Scores on the training data itself are shown only as a health check, never as a result.
02
A range on every number
Every headline number comes with its range of uncertainty and the count of trades behind it, so a real effect can be told from luck.
03
Rules that cannot be revised
When to enter, when to exit, how long to hold and how much to stake are all locked before any signal goes out. Changing a rule starts a new experiment; it never edits an old one.
04
The grader is code
Grading runs unattended each morning and publishes what it says. There is no editorial step between the result and the page.
05
Access
The sequence
V
Access opens in the order the evidence arrives.
The writing is open today. Access to the trades themselves comes later, once the models have earned it.
Available today
Free, forever
The writing & the record
The essays and The Graded Tape, public and free. Read the misses before you pay for anything.
Free
Tier 1
The letters
The full build in public: each experiment written up before it runs, what went wrong when one fails, and the reasoning behind every verdict. This is what funds the lab.
$20 / mo · $200 / yrComing soon
Tier 2 · limited seats
Trade-along
The system's signals, published to members after a stated delay, under rules fixed in advance. Capped by design; an edge stops working once everyone has it.
TBD / mo
App & API access
The dashboard
Built only after the models have proven out in the open. Planned for 2027.
Planned 2027Close
V
Read the misses first.
Start with the record and the failures. If the honesty holds up, the letters are where the work happens.