AI Stock Trading Challenge: 7-Day Performance vs S&P 500

•

 17 min video

•

 7 min read

YouTube video ID: eg_1NXDcoPk

Source: YouTube video by Nate Herk | AI Automation — Watch original video

PDF

This article details a seven-day stock trading challenge where an AI, specifically Codex with GPT-6 Astra, was tasked with managing a $10,000 portfolio. The primary goal was to outperform the S&P 500.

The Challenge Setup

The experiment involved giving Codex $10,000 of real money to trade stocks. Unlike a previous experiment, this time the user could make up to two changes daily to the AI's strategy. The ultimate success metric was beating the S&P 500; failure to do so would result in giving away a free VIP ticket to an upcoming event on October 17th and 18th.

Day Zero: Strategy Formulation

On day zero, before the market opened, the AI was tasked with extensive research to formulate a trading strategy. It delegated this to various "trading agents" who helped create a plan. The strategy involved six "wake-up" actions for trading and two for notifications. These actions were scheduled throughout the trading day:

  • Before market open: Read news, check the account, and choose stocks to watch.
  • One hour in: Look for the first qualifying trade.
  • Midday: Review positions and consider the last new trade.
  • Afternoon: Manage existing positions.
  • 2:15 PM: Close remaining positions.
  • Before market close: Confirm being out of positions and record the day's results.

A crucial aspect of the design was "continuity," where each agent would leave a "handoff message" for the next, ensuring a seamless workflow as if a single agent was managing everything. The routines were set to run in a single thread, allowing for easy monitoring. The system was powered by Codex with GPT-6 Astra on high settings.

Day One: Initial Setback

The first official trading day concluded with a slight loss, bringing the total to $9,991, a decrease of about $9. The user expressed frustration, noting that for a seven-day challenge, this lack of progress was concerning. The AI had bought Tesla stock with a stop-loss order, which was executed quickly when the stock dropped, risking over $3,000.

In response, the user made a tweak to the AI's strategy, emphasizing the urgency of the seven-day challenge. The AI acknowledged this, activating a "more active plan" with three entry reviews and less restrictive filters. The reward-to-risk ratio was lowered from 2:1 to 1.5:1, hoping for more aggressive trading on subsequent days.

Day Two: Technical Hiccup and Stagnation

Day two revealed a significant technical hurdle. GPT-6 Astra, due to its internal restrictions, would not execute trades automatically. It would strategize but not place orders. To circumvent this, a "GrokBot" named "Trader" was set up. Astra would send an email with trade recommendations to Trader, which then, via a webhook, would process the trades through Alpaca. Astra remained the "mastermind" for strategy, delegating execution to GrokBot.

Despite this fix, day two saw minimal activity, with the portfolio value changing by only $0.21. The user found this unacceptable, given only five days remained. The plan was to make the AI more active and aggressive by having it wake up more often (perhaps every 30 minutes) and reducing some of its "barriers" or risk-averse gates that prevented trades.

Day Three: Modest Gains and S&P Comparison

The update for day three (which occurred on a Sunday, so the market was closed) showed a modest gain. The portfolio reached $10,018, up $18. The current holdings included almost $2,000 in HPE, $1,200 in Exxon Mobil, and about $1,000 in Apple. Approximately $4,000 was actively in the market, with $5,800 still in cash.

The user planned to instruct the AI to become even more active and leverage more of the cash. A key point was reiterated: the goal wasn't to turn $10,000 into $15,000 in seven days, but to beat the S&P 500. At this point, if $10,000 had been invested in the S&P 500 on the challenge's start date (September 9th), it would have been down by about $5. The AI's portfolio, being up $18, was currently outperforming the S&P.

Day Four: Increased Activity, Significant Loss

Day four began with a morning update showing a decline to $9,876 overnight. However, a positive development was that 80% of the portfolio, approximately $8,000, was now actively invested, with only $2,000 in cash. This indicated the AI was becoming more aggressive as intended.

By the end of day four, the portfolio had dropped significantly, losing $225, bringing the total to $9,800. The daily change was down 2.26%, while the S&P 500 was down only 0.5%. This meant the AI had performed much worse than the S&P on this day. Holdings included CrowdStrike and an S&P ETF (which the user found interesting, given the goal to beat the S&P). With only three trading days left, the user planned further adjustments to the AI's strategy.

Day Five: Aggression and a Win Against S&P (Daily)

On the morning of day five, the AI's aggression was evident, with only $8 in cash and nearly $10,000 invested. The portfolio stood at $9,783, showing a daily loss of $13 at that moment.

By the end of day five, the portfolio had recovered slightly to $9,832.58, marking a gain of $36 (0.37%) for the day. Crucially, the S&P 500 was down 0.45% on the same day, meaning the AI successfully beat the S&P for the day. However, overall, the portfolio was still behind the S&P 500 by about $100 since the start of the challenge. The remaining cash was $63, with the rest invested. The user noted that short-term stock market movements are often small unless engaging in high-risk options trading, and expressed interest in continuing the experiment for a longer duration (30, 60, 90 days, or a year).

Day Six: Federal Reserve Impact and Further Decline

Day six brought an unexpected market event: the Federal Reserve unanimously voted to raise interest rates by 25 basis points (from 3.75% to 4%). This typically makes borrowing more expensive and can lead to lower stock valuations. The S&P 500 dropped immediately after the announcement, and so did the AI's portfolio.

By the end of day six, the portfolio was at $9,759, down $85 (0.87%) for the day. The S&P 500 was down 0.45%. The AI had sold a significant portion of its holdings after the Fed announcement, resulting in $3,600 in cash. It later picked up 80 shares of FPS. Overall, the AI's portfolio was down about $240, while the S&P 500 would have been down about $140, meaning the AI was still underperforming. The user planned to instruct Astra to invest the remaining cash for the final day.

Day Seven: Final Push and Challenge Conclusion

On the final day, the portfolio initially showed a significant surge, reaching $10,012, but then quickly dropped. The AI managed to sell a position before a further tank. With only $100 in cash, most of the funds were actively invested. The portfolio was up $153 on the day at one point.

At the close of day seven, the portfolio finished at $9,900, down $100 from the initial $10,000. For the day, the portfolio was up $182 (1.87%), outperforming the S&P 500, which was up 1.14%. However, over the entire seven-day challenge, the S&P 500 would have resulted in $9,980, beating the AI's performance by about 0.8%.

Conclusion and Future Plans

The AI ultimately lost the challenge, failing to beat the S&P 500. As promised, a free VIP pass to AIS Live (October 17th and 18th) would be given away to a commenter in the user's free community.

The user reflected that the first two days were too passive, and the AI only truly became active from day three onwards. The challenge only involved trading stocks, not higher-risk options, which could have led to larger gains or losses. The user expressed interest in continuing the experiment for a longer period (30, 60, 90 days, or a full year) to better assess the AI's viability as a trading strategy, as seven days was deemed insufficient to draw definitive conclusions.

  Takeaways

  • The AI, using Codex with GPT‑6 Astra, managed a real $10,000 portfolio over seven trading days, aiming to beat the S&P 500 benchmark.
  • Early in the challenge the AI was too passive, resulting in minimal activity and a small loss, prompting the user to adjust risk parameters and increase trading frequency.
  • By day three the portfolio showed modest gains and briefly outperformed the S&P 500, but aggressive adjustments on later days led to larger losses, especially after the Fed’s rate‑hike announcement.
  • On the final day the AI posted a daily gain that exceeded the S&P 500, yet the cumulative seven‑day return of $9,900 fell short of the index’s $9,980, resulting in a failed challenge.
  • The experiment highlighted the need for longer testing periods and possibly higher‑risk instruments to fully evaluate AI‑driven stock‑trading strategies.

Frequently Asked Questions

Why did the AI’s performance worsen after the Federal Reserve rate‑hike?

The Fed raised rates by 25 basis points, increasing borrowing costs and pressuring stock valuations; the AI responded by selling many holdings, moving to cash, which reduced exposure to the market’s rebound and caused its portfolio to decline more than the S&P 500 that day.

What role did the 'GrokBot' named Trader serve in the AI’s trading workflow?

GrokBot Trader acted as the execution layer; Astra generated trade recommendations and emailed them to Trader, which then used a webhook to place orders through Alpaca, allowing the AI to bypass its internal restriction against automatic trading.

Who is Nate Herk | AI Automation on YouTube?

Nate Herk | AI Automation is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.

Does this page include the full transcript of the video?

Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.

Helpful resources related to this video

If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.

Links may be affiliate links. We only include resources that are genuinely relevant to the topic.

Full transcript is not shown on this page

This page focuses on the summary and original notes. For full verification, refer to the original YouTube video.

PDF