I built a nine-feed trading pipeline — then shut it down.
A multi-agent crypto and prediction-market signal system: nine data feeds, a convergence scorer, whale-alert and contrarian channels, autonomous paper-trading. I measured a losing edge honestly — 81 trades, ~26% win rate, -$30.55 — killed the over-engineered version, tried a leaner one, and wound the whole thing down. Zero real dollars risked.
Could a fleet of agents find a real edge?
The fleet’s stated objective was “make money.” Markets are the obvious place to test it.
The hypothesis, borrowed from a third-party course claiming an “AI contrarian” strategy, was that agents could bet against the crowd.
Prediction and crypto markets are where you’d test whether AI’s real advantage — reading and synthesizing huge amounts of data faster than a human — translates into a tradeable edge.
The whole point of the experiment was to find out if that edge actually existed for us, on live markets, with our own pipeline — before ever committing real capital.
Whose numbers are whose.
Two very different figures get confused here. This page keeps them apart on purpose.
“~11,000% ROI” belonged to the course.
- The eye-catching ~11,000% ROI figure came from the third-party course that inspired the experiment.
- It was a marketing claim about someone else’s strategy — never a result my system produced.
Do not read the course’s advertised return as this project’s outcome. It is shown only to explain where the hypothesis came from.
A small net loss on paper.
- My system’s actual, measured outcome: -$30.55 across 81 closed paper trades.
- 25.93% win rate. A losing edge, measured honestly.
- Zero real dollars were ever at risk — paper-traded throughout.
Build it properly, then audit it honestly.
V1 was the maximalist version. The decisive move was the honest audit that followed.
V1 was a nine-feed convergence pipeline feeding a scorer, with autonomous paper-traders auto-executing on high-convergence signals, whale-tracking, and a dashboard. It ran, it traded on paper, and it lost — and its intelligence layer degraded to garbage, with 5 of 9 feeds failing and signals coming back null.
Rather than tweak endlessly, I had the agent lay out the full profit-and-loss by category. So I killed V1 entirely — deleted the crons, purged 210 files — then tried a leaner research-only V2, and eventually stood that down too.
Over-engineered and under-performing — complex architecture producing zero intelligence. So I killed it, rather than polish a losing system.
— The kill diagnosis · from the auditThe V1 pipeline.
Nine feeds converge on a scorer; only multi-source agreement flags a trade. High-convergence signals auto-execute on paper-traders; whale and contrarian signals fan out to alert channels; everything syncs to a dashboard.
Nine feeds fan into a convergence scorer that only flags a trade when multiple independent sources agree. High-convergence signals route to paper-traders — no real capital — while whale and contrarian signals fan out to alert channels. Everything syncs to a dashboard on Postgres and Convex, with a backtester replaying trades against new rules. On the final run, 5 of 9 feeds were failing and the scorer produced zero tradeable signals.
Agents
Data Feeds (9)
Signal Layer
Execution
Data / Infra
I/O
The audit, unvarnished.
Feb–Mar 2026 · verified from the honest P&L audit · paper trades throughout.
The only positive category was essentially coin-flip binary markets where a few big bets got lucky; everything else was flat or negative. The infrastructure was built out well beyond what it ever produced — the alert channels carried a handful of messages over about a day, then went quiet.
The valuable outcome was the decision to stop.
I built a genuinely sophisticated system, ran it on paper, and it did not find an edge. The portfolio value isn’t a return — it’s the discipline.
Measured it. Called it. Cut it.
- Forced a real, category-level P&L audit instead of more feature work.
- Accepted a losing verdict on a system I’d invested real build effort in.
- Killed and purged the over-engineered V1 — 210 files removed so it couldn’t rot in the workspace.
- Tried a leaner V2 hypothesis (research-driven, conviction-sized, logged reasoning).
- Wound V2 down too when it didn’t earn its keep — crons removed, trading group cleared Mar 22.
Framed exactly as what it was.
- This is a started-then-abandoned experiment — not a profit, not a running system.
- It shows range: multi-agent architecture, market-data engineering, backtesting, dashboards.
- More importantly, it shows the judgment to stop — most people keep polishing a losing system.
No real capital was ever exposed. Do not imply this made money or is still running. Its value is breadth plus the discipline to walk away.
From maximalist to wound down.
How the experiment ended.
Nine-feed convergence pipeline, autonomous paper-traders auto-executing on high-convergence signals, whale-tracking, and a synced dashboard.
It traded on paper and lost; 5 of 9 feeds failed and the scorer produced 34 neutral, null signals — zero tradeable output.
A category-level audit exposed the losing edge (-$30.55, ~26%). Diagnosis: over-engineered and under-performing. V1 crons deleted, 210 files purged.
Inverted the thesis: no signal-spam, one $10K paper account, conviction sizing, max 3 trades/day, capture the “why” behind every trade.
V2 research crons removed on strategy shutdown; the trading group was cleared out. No real capital was ever exposed.
I owned the hypothesis, the risk controls, and the kill.
I scoped the experiment, directed the build, set the safety posture — paper-trading only, no real money ever went in — and forced the honest audit.
When performance was ambiguous I demanded the category-level P&L rather than more feature work, made the call to kill V1 and permanently purge it, defined the leaner V2 thesis, and decided to wind the whole thing down. The agents built and traded; I owned the hypothesis, the risk controls, the audit, and the kill decision.