Work · Trading
Journal, coach & backtest lab
Two systems for same-day options trading: a journal with an embedded AI coach that knows your actual trades, and a research pipeline whose job is to tell the truth about whether an edge exists — especially when the answer is no.
Trading journal
Dalton: a coach grounded in your own data
Dalton is an AI trading coach in the Market Profile tradition, embedded directly in the trade journal. Before every reply it pulls today's trades, recent journal entries, and the market context from the database — so the coaching is about what actually happened, not what you pasted in. Upload a chart screenshot and a vision model reads it into the conversation.
It runs two modes: fast feedback during a live session, and a structured end-of-day recap that saves straight into the journal. The coaching framework separates read errors (you misjudged the market) from process errors (you broke your own rules) — because they need different fixes.
vision
reads uploaded chart screenshots
live SQL
context injected from the trade database
multi-model
routed across providers with streaming
0DTE backtest
Turn instinct into statistics
A staged Python pipeline tests intraday option entries against real historical tick data: signal generation, strike discovery that targets a premium band by session phase, and outcome measurement on one-second option quotes. A fair-value gauge prices each option with Black-Scholes using the day's volatility index, so the system knows when it's overpaying.
14,268
per-contract quote files downloaded
5,271
entries measured across 17 months
$40
hard spend ceiling per research run
The honesty centerpiece
We paid $24.75 to learn our edge was fake
The naive signal loses
The obvious strategy — chase the confirmed indicator cross — tested decisively negative: −11.3% per trade (t = −5.98) across 2,657 liquid entries. That result retired the idea before it cost real money.
The scorer that didn't generalize
A machine-learning entry scorer looked promising until blocked time-series cross-validation and a permutation test (p = 0.553) showed no real signal. Published as a negative result, not quietly dropped.
The bug ledger
A subtle bar-labeling bug was leaking about one minute of hindsight into every simulated entry. The fix is documented in a numbered ledger — and every statistic computed before it was declared invalid and purged.
What survived
One exit-management rule cleared a strict out-of-sample statistical gate — and the writeup itself flags that its profits are concentrated in the fattest tail. The caveat ships with the result.
Say hello
Curious about any of this?
No pitch — the work is the pitch. If something here sparked an idea, or you'd like a closer look at how a piece of it fits together, I'd genuinely love to talk about it.
Say hello