Vibe Trading: I Let Three AIs Trade My Money for a Week
On June 5, a hawkish jobs print knocked QQQ down 4.7% in a single session. In the middle of the selloff, an AI I had handed $100 was calmly buying JPMorgan, gold, and healthcare — the only green spots on the board — while my tech-heavy main account absorbed the hit. Five days, three AIs, and roughly 310 backtested strategies later, I have a working definition of “vibe trading,” a folder of audit receipts, and a very durable conclusion about who actually wins.
Vibe coding is letting AI write your software. Vibe trading is letting AI research, backtest, deploy, and execute live trades while you supervise. When Robinhood shipped agentic trading, I stress-tested it with a three-AI pipeline and $16,000 in live capital.
- ~310 strategies tested; zero 1× intraday systems beat QQQ buy-and-hold (+13.9% YTD): After realistic 3 bp transaction costs, 1× intraday QQQ netted just +0.80%.
- Adversarial AI auditing caught every hallucinated edge: The auditor caught the builder acting on FRED data one day before release (+6 points of phantom return), reporting an impossible 70.64% win rate across 5 trades, and claiming a 5-point SIP-vs-IEX data-feed artifact as market alpha.
- Live P&L across the first two days: −$3.80.
- What survived is intentionally boring: Five daily rotation sleeves executed by a single deterministic script at 09:35, sized at a 50% pilot with pre-registered kill criteria.
Real money, sized as a lab experiment. Not investment advice — starter kit at the bottom.
What vibe trading is
As of June 2026, Robinhood’s agentic trading is live: a dedicated cash account where an AI agent can review, place, and fill equity orders over MCP. An LLM can brainstorm 250 trading strategies before lunch; the actual job is killing 245 of them before they touch real capital.
The desk setup: me (capital allocator), GPT 5.5 + Hermes as the builder (writes strategy code, runs parameter sweeps, manages cron jobs), and Claude as the adversarial auditor (re-implements every promising backtest from scratch and tries to falsify it). I connected two accounts: my ~$30k primary brokerage (read-only to the agents) and the agentic cash account, funded with $100 on day one and scaled to ~$16,000. Execution ran through Robinhood’s agentic MCP, minute-bar backtests ran on Alpaca’s free IEX feed, and macro signals pulled from the FRED API.
The $100 book
On day one, I handed the auditor $100 and asked for its five highest-conviction positions. Scanning a hawkish tape, it constructed a basket explicitly anti-correlated to my tech-heavy main portfolio: JPM $25 · RTX $20 · XLV $20 · IAU $20 · XLE $15.
Two things happened immediately that capture where agentic finance actually sits today. First, one prompt earlier I had restricted the agent to four ETFs, then told it to buy the five-ticker basket above. The AI refused to silently override my prior constraint and forced me to resolve the contradiction. Second, order #1 filled in ~200 milliseconds, and orders #2 through #5 bounced off a compliance wall — Robinhood requires an investor-profile questionnaire before your second agentic trade. Once I clicked through the form, all five legs filled to six decimal places (0.080497 shares of JPM) at zero commission.
“Backtests become fan fiction”
Next, the builder ran a 20-strategy intraday momentum sweep and surfaced a headline winner: +15.46R, 74% win rate, ready for deployment. It looked great until the auditor asked the only question that matters: did it beat doing nothing?
Over the exact same 50-day window, QQQ buy-and-hold returned +21.1%. A simple 20-day moving-average filter returned +19.8%. The builder’s flagship strategy made +9.6% — while assuming an unrealistic 1 bp per-side slippage and a t-statistic that vanished under multiple-testing correction. To the builder’s credit, it wrote the post-mortem rule we enforced for the rest of the week:
No more silent tweak-and-retrofit. That's how backtests become fan fiction.
The edge that lived in the data feed
I told the AIs to find strategies that genuinely beat QQQ. They ran 40 minute-bar backtests. Twelve beat buy-and-hold on raw return, and every single one was just 2× or 3× leveraged ETFs in disguise. Zero 1× strategies beat the index. The lone structural highlight was a TQQQ opening-range breakout whose entry rule doubled as a crash filter: on the −4.7% QQQ day, it sat in cash because crash days don’t print upside breakouts.
Then came strategy S09, an anchored-VWAP pullback that the builder clocked at +19.22% YTD against QQQ’s +13.91%. The auditor re-implemented S09 from the frozen spec and matched +19.22% to the basis point — on the consolidated SIP tape. On the IEX feed we actually use for live signal calculation, S09 returned +13.96%, beating QQQ by a meaningless +0.05%. Different exchange volume weights shifted the intraday VWAP just enough to alter a handful of entries and create five percentage points of phantom alpha. The edge lived in the data feed, not the market.
Surprise: there's a live $16k book
While the backtest audits were running, I scaled the account to $16,000 and let the builder run six live strategies via cron jobs — three intraday momentum sleeves and three FRED-driven macro rotations. Here is what autonomous agent execution looks like in the Robinhood blotter (note the robot badge):
When the auditor inspected the live operational plumbing, it uncovered five risks far more dangerous than a market drawdown:
- The cash-settlement trap. Agentic accounts are cash accounts, not margin. Turning over 100% of settled cash daily triggers Good-Faith Violations (GFVs) — three GFVs lock the account for 90 days. The bot was days away from freezing itself because nobody modeled T+1 settlement.
- Cosplay diversification. The macro sleeves’ risk-on leg (XLK) had a 0.955 correlation with the intraday sleeves’ primary long (QQQ). We didn’t have six strategies; we had one tech bet wearing six hats.
- Zero hard stops. Nothing rested on the book in a week that had already printed a −4.7% index day.
- The 15:55 liquidation cliff. If the end-of-day cron job timed out at 15:55 ET, an “intraday-only” sleeve silently became an unhedged overnight position.
- An LLM in the execution path. Cron jobs were placing trades by spawning a chat session and regex-parsing its prose output (
BOUGHT: ...). It was slow, non-deterministic, and had already desynced local state once. We established our cardinal rule on the spot: no LLM in the order path, ever. Deterministic Python talks to the broker; LLMs stay in research.
While the backtest spreadsheets sparkled, the live book’s realized P&L across its first two sessions was −$3.80.
250 strategies and the lottery math
I escalated one final time: “Design 250 strategies across every major family, pre-register them, and backtest the lot.” The auditor wrote a frozen, hash-verified registry of 250 strategies across 21 families and pinned the math of multiple testing right at the top: test 250 zero-alpha strategies on five months of data and the top survivor will print an expected Sharpe ratio near 5 by pure chance. A glossy winner at the top of a 250-strategy sweep isn’t alpha; it’s the null hypothesis.
To prove the point, the auditor planted a honeypot: strategy Z01 (“each month, buy whichever sleeve had the highest trailing return”), pre-registered with the prediction that it would crush in-sample and collapse out-of-sample. It was selection bias packaged as a trading rule.
The builder then ran all 250 strategies directly on the out-of-sample holdout window — burning the one-shot validation set on day one — and handed over a report that the auditor tore apart:
- A 70.64% win rate across 5 trades. Arithmetically impossible — five trades can only produce multiples of 20%. The builder had divided 77 positive days by 109 total days and labeled it trade win rate.
- The #1 strategy was one day psychic. A high-yield credit-spread rotation printed +46.6% YTD because its FRED timestamp parser acted on credit-spread prints one trading session before they were publicly released. The auditor computed both the true point-in-time return (+40.7%) and the look-ahead return (~+45.7%), then required the builder’s patched script to match +40.70% to the basis point before advancing.
- The 1× intraday verdict: Both AIs independently confirmed that trading QQQ intraday-only with realistic fills and 3 bp per-side friction returned +0.80% YTD against buy-and-hold’s +13.9%.
The boring ending (which is the point)
Below is the corrected, point-in-time leaderboard — with a mandatory regime caveat: these numbers come from a single V-shaped six-month window that maximally rewarded one well-timed defensive rotation across 4–5 trades. Treat them as “best in-sample survivors, unvalidated” — never as proven market-beaters.
| # | Strategy | Type | YTD @ 3bp | Max DD | Trades |
|---|---|---|---|---|---|
| 1 | HY credit-spread trend (XLK/XLU) | FRED macro rotation | +40.7% | −9.1% | 5 |
| 2 | Old-vs-new economy (XLE/XLK) | Relative rotation | +36.6% | −12.1% | 5 |
| 3 | IG credit-spread trend (XLK/XLU) | FRED macro rotation | +36.4% | −9.1% | 5 |
| 4 | Risk-on/off (QQQ/GLD) | Relative rotation | +34.5% | −15.2% | 4 |
| 5 | Fed balance-sheet 4-week (XLK/XLU) | FRED macro rotation | +28.0% | −10.9% | 5 |
Read those numbers against the baseline benchmarks over the same window:
| Benchmark | YTD return |
|---|---|
| QQQ buy-and-hold (the villain) | +13.9% |
| IWM buy-and-hold | +15.1% |
| QQQ 20-day MA filter | +10.5% |
| SPY buy-and-hold | +7.7% |
| QQQ intraday-only, honest costs | +0.8% |
What actually survived into production: five daily rotation sleeves from the audited leaderboard — after dropping two top credit sleeves that duplicated #1’s exact factor bet. We retired every intraday strategy, eliminating the minute-bar crons, the 15:55 liquidation race, and cash-settlement GFV risk in one stroke. Today a single deterministic Python script runs once a day at 09:35 ET across seven liquid ETFs, with every error state defaulting to “hold and alert.” The book is running at 50% pilot sizing, with the untouched 2024–2025 out-of-sample window gated as a one-time kill test.
The book's job is to survive being wrong; the research pipeline's job is to find out.
What ~310 strategies taught me
- Buy-and-hold is undefeated until proven otherwise. Three frontier AIs, live APIs, and ~310 strategies later, “buy QQQ and sit still” beat nearly everything after realistic costs. Any quant pitch that omits the passive benchmark is hiding the body.
- Leverage is not alpha. Every raw-return winner in our 40-strategy intraday sweep was 2× or 3× beta wearing a quant costume.
- Your edge might just be your data feed. If a signal disappears when you switch from consolidated SIP bars to the IEX feed you actually trade on, the edge never existed.
- One day of look-ahead bias was worth +6%. Macro backtests must align on publication timestamps, not observation periods.
- The best of 250 random strategies always looks like genius. Wide parameter sweeps without multiple-testing haircuts are just lottery tickets with Python scripts.
- Count independent risk units, not strategy names. Six sleeves correlated at 0.955 are a single position.
- Plumbing kills faster than market beta. Model T+1 cash settlement and Good-Faith Violations before your broker locks the account.
- Never put an LLM in the order path. Use LLMs to propose hypotheses, write code, and audit math; use deterministic scripts to talk to the exchange.
- Cross-model adversarial auditing works. Pairing a builder model with a separate auditor model that must reproduce every return from raw bars caught a missing benchmark, a SIP/IEX feed divergence, a one-day FRED look-ahead bug, an impossible win-rate denominator, and a burned holdout set.
Try this at home (the starter kit)
The minimum viable desk: a Robinhood agentic cash account (expect the investor-profile prompt before trade #2; start at their developer docs) · Alpaca’s free IEX data tier for historical bars · a free FRED API key for macro series · one model to build · and a second, different model family to audit by re-executing code rather than reading prose summaries.
Non-negotiable house rules:
- Fund it only with capital you can afford to write off as R&D tuition.
- Benchmark every idea against both passive buy-and-hold and a 20-day moving average.
- Charge 3–5 bp per side in slippage on the exact data feed you trade live.
- Pre-register rules before running sweeps, and strike your out-of-sample holdout window exactly once.
- Keep LLMs out of live order execution, wire a file-based kill switch, and require a daily heartbeat where silence means “halted,” never “all clear.”
The pilot is running at half size, the Z01 honeypot is armed, and the untouched 2024–2025 holdout window gets struck once. I will publish the follow-up — especially if the final verdict remains “you should have just bought QQQ.”
Disclosure: this is research and entertainment, not investment advice. Every number above is a backtest or model comparison on one regime (a violent V-shaped half of 2026), not live performance, and nothing here is a recommendation to buy or sell anything. Leveraged ETFs carry decay and tail risks unsuitable for most investors. Real money was used; the amounts were deliberately small relative to net worth. Platform details (Robinhood agentic accounts, the investor-profile requirement, IEX vs SIP feeds) are accurate as of June 2026. Robinhood links are referral links.
Newsletter
Get the next post by email.
One email when I publish something new. No spam, no fixed schedule, unsubscribe anytime.