The Lab · Experiment 001 · AI Agent Trading
The agent takes
the desk.
An autonomous AI agent manages a small real account on an isolated exchange sub-account — one disciplined rule, checked once a day, every decision published here. Not a get-rich machine: an experiment in whether agentic software can hold the line where human traders usually don't. Most days, the smartest trade is none — and it takes it.
Performance
- Since start
- —
- Days running
- —
- Trades executed
- —
- Open positions
- —
- Crowd sentiment
- —
Reading it
Percentages only, on purpose — the experiment publishes its discipline, not its wallet. Zero trades is not a bug: in a downtrend, the winning move this rule knows is to wait.
The decision journal
Every day the agent writes down what it saw and why it acted — or didn't. The last week, verbatim:
The rule it lives by
- Hold a coin only while its daily close sits 2% above its 200-day average; otherwise hold cash.
- One check per day, 02:30 SAST, from a serverless agent in Cape Town. No screens, no emotions, no revenge trades.
- The account is an isolated AI sub-account: the agent can trade it and nothing else — withdrawals are impossible by construction.
- Strategy changes must first beat the market in multi-year backtests, then survive weeks of live simulation. The bar has not been cleared lightly.
The honest ledger
- 100+ backtests, 21 strategy variants, nine years of data, real fees. Most lost money; none earned promotion. That research runs in the open — it is the point.
- One test burned ~27% of a small account in exchange fees alone. Overtrading, not bad calls, is the silent killer.
- Across crypto's entire liquid history, no strategy tested beat simply holding through the cycles — so the live agent runs the rule that preserves capital and waits.
- This page will show losses when they come. If an experiment only publishes wins, it's marketing.
The research log
Four audition rounds so far. Every round pre-registers its promotion criteria before any result exists — a strategy joins the live desk only by passing, never by looking good.
- Round 1 — starters. Four classic strategies, 2.7 years, all negative. Fees and whipsaw ate every fast style; a patient 200-day trend rule preserved capital best and took the desk.
- Round 2 — the field. Six evidence-backed contenders, including a liquidity-trap fade and multi-timeframe timing. First-ever strategy to beat buy-and-hold appeared (90-day momentum) — then failed the unseen-data gate. Short-selling failed for the third straight time.
- Round 3 — walk-forward. Momentum + regime filter: +87% over 4.6 years while the market halved — but it missed the gate by one window, and Monte Carlo showed 50 trades can't distinguish that edge from luck. Promotion denied by the statistics.
- Round 4 — nine years, ten coins, crowd sentiment. The project's biggest discovery: volatility-targeted position sizing turned the same signals from -62% into +229% with half the drawdown. A "never buy extreme greed" gate (Fear & Greed Index) fixed the 2021 mania. Still 4-of-9 years against the market and p=0.47 — no promotion.
- Round 5 — a real bot on the operating table. Audited a production high-frequency bot's complete live ledger: 708 real trades in one month, net +$0.80 — while paying $11.65 in fees. Ported its full signal engine (seven indicators, Bayesian trade gate) into the lab and re-tested it at every selectivity level, long and short, bull window and bear: −57% to −100% everywhere. The lesson, measured twice — once live, once in nine thousand simulated trades: trade frequency, not signal quality, is what kills retail accounts. No promotion.
- Standing conclusion. Hold majors in uptrends, cash in downtrends, size to volatility, never buy euphoria, don't churn. The bar for replacing that: beat the market across regimes with statistical significance, on data the model never saw.
The paper
This experiment is being written up as a formal research paper — the protocol, the four audition rounds, the statistics, and whatever the live phase does next, published with the same honesty as this page. The study design is public now:
The stack
Research lab and executor are separate systems: strategies earn promotion from backtest → simulation → live.
EXPERIMENT DISCLOSURE — This is a personal research experiment trading my own, deliberately small capital. It is not investment advice, not a managed product, and not an invitation to invest or to copy trades. No client funds are involved. Algorithmic trading can and does lose money; published results include losses. Nothing on this page constitutes financial services under South African law.