AI Creators
Nate Herk
I Gave GPT 6 Astra $10,000 to Trade Stocks...And This Happened
Sep 28, 202618 min watch
Quick Overview
In this video, Nate Herk documents a seven-day automated investing experiment using real capital. The project tests whether an AI-driven workflow built on GPT 6 Astra and Codex can successfully manage a stock portfolio and beat the S&P 500 index.
Key Points
- 1.Nate Herk allocated ten thousand dollars of real capital to test an automated stock trading system driven by GPT 6 Astra over a seven-day period.
- 2.Because of safety restrictions preventing direct automated trade execution inside GPT 6 Astra, trades were passed via email webhooks to a secondary bot connected to Alpaca.
- 3.The initial configuration proved overly conservative, requiring rule adjustments such as lowering the minimum reward-to-risk ratio from 2:1 to 1.5:1.
- 4.A 25 basis point interest rate increase by the Federal Reserve during day six led to broader market drops and triggered portfolio sell-offs.
- 5.The AI trading experiment ended with a portfolio balance of $9,900.39, finishing approximately 0.8 percent behind the benchmark S&P 500.
Summary
Nate Herk begins a seven-day experiment giving GPT 6 Astra $10,000 in cash to trade live equities, setting the benchmark performance target as the S&P 500 index. Unlike a previous challenge where the setup could not be touched once started, the rules here allow up to two manual strategy adjustments per day. The automated setup uses Codex to coordinate six daily scheduled routines that analyze news, select watchlist tickers, enter positions, manage risk, and close trades before the market bell. Because native execution protections in GPT 6 Astra reject direct order submission, Herk routes trade signals via email webhooks to a secondary bot connected to the Alpaca brokerage platform.
On the initial trading day, the agent executes only a small trade in Tesla that triggers an immediate stop-loss, leaving the portfolio down approximately $9 at $9,991.40. Day two shows similarly stagnant performance, recording a daily gain of just 21 cents due to overly restrictive trade filters. In response, Herk modifies the prompt parameters to lower the required reward-to-risk ratio from 2:1 down to 1.5:1 and enables more frequent scanning intervals.
By day three, the updated parameters generate higher trading activity, establishing positions in Hewlett Packard Enterprise, ExxonMobil, and Apple, which pushes the portfolio balance up to $10,018.47. On day four, the bot rotates into broader index and sector ETFs including SPY, CrowdStrike, and XLV, putting roughly 80 percent of the capital actively to work. A downturn in holdings like CrowdStrike causes the account balance to pull back to $9,792.65 by market close, trailing the benchmark index on the day.
Day five sees the account recover slightly to $9,832.58, outperforming the S&P 500 for that individual session. However, on day six, the Federal Reserve announces a 25 basis point benchmark interest rate hike. The resulting macroeconomic volatility sends the major indices into a steep drop, prompting the AI system to liquidate several positions into cash and ending the session down at $9,750.31.
On the final day, the agent takes positions in sector funds such as XLK, Marvell Technology, and XLV, capturing a daily rise of $182.19 or 1.87 percent, which beats the S&P 500 daily return of 1.14 percent. Despite this strong finish, the account ends at $9,900.39, representing an overall loss of 1.00 percent. A passive buy-and-hold SPY strategy over the exact same time frame would have ended at $9,980.63, a 0.19 percent loss, meaning the AI system fell short of beating the benchmark by approximately 0.81 percentage points.
Experiment Setup and Automated Architecture
Nate Herk establishes a seven-day live trading challenge using $10,000 in real capital, aiming to outperform the S&P 500 index. The automated pipeline relies on GPT 6 Astra running inside Codex to conduct market research and schedule routines across the trading day. Due to platform restrictions preventing the model from executing orders directly, an intermediary Grok bot named Trader is configured to receive trading instructions via email webhooks and execute the orders through the Alpaca brokerage API.
Early Trading Days and Strategy Adjustments
During the first two days, trading activity is minimal due to strict filtering parameters and a conservative 2:1 reward-to-risk threshold, resulting in a negligible change of negative $9 on day one and positive 21 cents on day two. To increase market exposure, the strategy is adjusted to a 1.5:1 reward-to-risk ratio and more frequent check-ins. By day three and day four, larger positions are opened in assets such as SPY, CrowdStrike, and XLV, bringing market exposure up to approximately 80 percent of the total portfolio.
Macroeconomic Shocks and Final Results
On day six, the Federal Reserve raises benchmark interest rates by 25 basis points, triggering a sharp decline across the broader stock market and prompting the automated agent to liquidate multiple positions into cash. Despite a rebound on day seven where daily gains reached $182.19, the challenge concludes with an ending account balance of $9,900.39. This represents a 1.00 percent loss compared to a 0.19 percent decline for a buy-and-hold SPY position over the same window.
The Bottom Line
The experiment demonstrates that while multi-agent workflows can research and place trades through external API integrations, short-term market noise and sudden macroeconomic policy shifts significantly impact automated strategies. The AI system ended the week with a 1.00 percent net loss, failing to outperform the benchmark index's 0.19 percent decline. The video leaves open whether longer time horizons or higher-volatility instruments like options would yield better relative performance under AI supervision.
FAQ
What is GPT 6 Astra and how is it used for trading stocks in this experiment?
GPT 6 Astra is an AI model used within Codex to automate stock market research, evaluate trading signals, and schedule daily routines to manage an investment portfolio.
Why could GPT 6 Astra not execute stock trades directly through the brokerage account?
Safety and model restrictions prevented GPT 6 Astra from executing financial transactions directly on behalf of the user, necessitating an email webhook handoff to a secondary bot connected to the Alpaca API.
What strategy adjustments were made after day one to increase trading activity in the challenge?
The minimum reward-to-risk ratio filter was reduced from 2:1 to 1.5:1, relative volume requirements were lowered, and review schedules were made more frequent to allow more candidate trades to pass.
How did the Federal Reserve interest rate announcement on day six affect the trading portfolio?
The Federal Reserve raised benchmark interest rates by 25 basis points, sparking market-wide drops that caused the portfolio to lose $85.73 on the day and trigger multiple asset liquidations into cash.
Did the GPT 6 Astra automated trading strategy beat the S&P 500 benchmark over seven days?
No, the AI trading system finished down 1.00 percent with an ending balance of $9,900.39, while the S&P 500 benchmark lost 0.19 percent over the same period.
Worth watching for
Individual investors, software developers, and automation enthusiasts interested in integrating large language models with algorithmic trading workflows and brokerage APIs.
- algorithmic-trading
- ai-agents
- gpt-6
- alpaca-api
- stock-market