Five parts, one loop
A conversation, a connection to the tools, an engine that runs experiments, the results it saves, and a place to inspect trades.
Codex
Let's test it across conditions: candle size, trend direction and how far ahead we measure. Each one is compared with ordinary candles.
What did the trend filter add, compared with candle size alone?
Tool request through MCP
- Varies
- 3 factors, 18 configurations
- Data
- Research period only
- Compared with
- Ordinary candles
Research Engine
Candle size
compared with the typical range
Trend
direction of the last 50 hours
Measured after
did price keep going?
Held constant One instrument, hourly candles, research period, the same baseline.
0 of 18 configurations
Saved Results
Kept going more often than ordinary candles About the same
Research period Validation period
configurations recorded in the trial ledger, including the weak ones.
Read through MCP
Saved results for the conversation
- Configurations
- Validated
- Computed in chat
- Nothing. The numbers come from the run.
Simulation, validation period
- Entry
- Next hour's open, in the candle's direction
- Stop
- One typical range away
- Exit
- After hours, or at the stop
- Costs
- per trade, a placeholder model
- Trades
- Average, after costs
- Positive
Trade Explorer
Synthetic demo
It starts with a question
A research idea, in plain words, in a Codex chat.
The plan goes out through MCP
MCP connects the chat to the research tools. The engine does the work.
Three conditions to vary
Candle size, trend and how far ahead to measure. Everything else stays fixed.
How a large candle is defined
A candle's range is compared with the average range of the previous 20 hourly candles. "2×" means at least twice that typical range. Candles count only when the body is at least 40% of the range, which gives each one a direction.
One idea becomes 18 experiments
3 sizes × 2 trend settings × 3 horizons, each run on the research period.
Every result is saved and comparable
Each cell is one configuration, measured against ordinary candles. Open a cell to see what it tested.
Why the cells are not independent
Every 2.5× candle is also a 2× and a 1.5× candle, and the three horizons reuse the same events. Neighbouring cells share evidence, so a pattern across many cells is not the same as many separate findings.
A separate period checks what held
The strongest cells are measured again on later data, chosen by a rule set before looking.
The rule, and the two periods
The results come back
Codex reads the saved run through MCP. Nothing is recomputed in the chat.
Explained in plain language
What stood out, how many events it rests on, and what held up.
The next question comes from the evidence
Each answer points back to the cells it rests on. Select one to open it.
One configuration, simulated as trades
Explicit entry, exit and cost rules turn the strongest validated cell into trades.
Inspect every trade
The trades in Explorer come from that same simulation. Step through them, losses included.
What I built
A research engine, a Codex integration through MCP, and Trade Explorer, so every step from question to trade can be traced.
One decision behind it. Saved runs are plain files and are never rewritten. The database is only an index, and can be rebuilt from them.
View the code on GitHubAbout this demo
All prices, results and trades here are synthetic, generated from one seeded run so the chat, the matrix and the trades agree. The page shows the intended workflow. Parts of it, such as comparing and explaining runs through AI, are still being built, and the demo will be checked against the product before launch.
What it does not do
No live trading, paper trading or broker execution. One instrument and one open position per experiment. The cost model is still under review, so simulated results are not estimates of real money.