I just gained access to the new and improved o1 model. Here’s how I’m using it to time the market
12 min readWith the stock market down over 5% in the past week, many new investors are asking themselves this question:
Should I sell?
When most people answer this question, they are speculating based on gut feel or recent experience. Seldom do people actually look at the data in order to inform their investing decisions.
This leads to rash decision-making, fear, uncertainty, doubt, and poorer profits.
Today, I’m going to show you how to use historical data to help inform your decision-making process. This article will give you the tools to help you decide if you should buy more, hold, or sell during certain market conditions.
Follow this guide to define a seven-calendar-day drop, inspect the subsequent drawdown and return, then translate that question into testable rules. The delivered result is a repeatable research method; the backtests below are historical examples.
Let’s get started!
How to Follow Along: a Google Collab for Investing
This article is designed to be interactive and give you the tools needed to make these types of decisions for yourself. Thus, I strongly encourage you to not just read, but follow along.
To do so, you’ll need to create a free NexusTrade account. NexusTrade is the platform we’ll use to perform the research to answer questions such as “should I sell during this pullback?”.
NexusTrade - AI-Powered Algorithmic Trading Platform
Learn to conquer the market by deploying no-code algorithmic trading strategies.
nexustrade.io
If you’re not able to create an account at this moment, not to worry! This article will contain plenty of screenshots and links that you can refer to.
One of the most important links is the direct link to the conversation I had with Aurora, the AI assistant that will help us perform this analysis. This direct link is more streamlined than this article and will remind you how to perform the analysis without you having to read this entire article.
SPY financial analysis and an SPXL trading strategy
SPXL is Direxion’s daily 3x S&P 500 ETF. Its target is three times one day’s index return before fees and expenses; it does not target three times the return over a longer holding period. SPY’s historical downside averages therefore cannot be transferred directly to SPXL. Read the issuer’s daily objective.
An in-depth analysis of SPY's performance, including instances of 5% drops, drawdowns over 180 and 365 days, and a…
nexustrade.io
This is vaguely similar to Google Collab, a platform coders use to share their scripts. With the conversation, you can continue from where I left off, copy the messages, and use it as a guide to perform your own analysis.
Finally, I encourage you to share this article with a friend to help me accomplish my mission of making data-backed systematic investing available to everybody, not just the “Wall Street Elite”.
Without further ado, let’s get started!
Experimental Setup: Our Choice of Language Model
This article will be comprised of three main objectives:
- Querying for financial data
- Performing financial analysis
- Running historical simulations
To do this, I will be using the brand new OpenAI O1 model. the O1 model is the successor to the O1-preview model. It support function calling, which allows us to generate synthetically valid objects — something useful for LLM applications.
The O1-model is a reasoning model, which means that it thinks before it answers complex questions. This is extremely useful for a financial application, where accuracy is paramount.
However, this thinking does have its drawbacks. For one, it's extremely slow, oftentimes taking several minutes to answer complex questions. It's also extremely expensive costing several times more than some of the other language models.
Within NexusTrade, the model we choose is configurable, meaning we could choose other models, depending on our use-case. This includes Sonnet 3.5 and Gemini Pro – two advanced language models from Anthropic and Google respectively.
Thus, we have to make trade-offs between speed, accuracy, and cost. For this article, to demonstrate the full capabilities of the stronger large language model, we are going to use the brand new O1 model.
To view the exact queries the model used, you can click the “View Stock Screener Debug Info” attached to the AI messages in the chat.
Using the O1 Model to Outperform the Market
By using the best language model, we can analyze the market with unprecedented speed and accuracy. Here’s how.
Within NexusTrade, the large language model is attached to a comprehensive financial database. The model acts as an interface between the database and the user. It transforms our user questions into queries, which are used to accurately answer our financial questions.
In this December 2024 example, an unconnected ChatGPT chat did not have the financial database and query tools used here. To inspect the same workflow, use a client connected to the required financial records and tools, such as NexusTrade. The difference is the available data and tools, rather than the chat interface alone.
Now, let’s get started with our analysis.
Performing financial analysis
We’re going to use Aurora to figure out what to do during a market downturn.
To start, I’m going to ask the following question:
Since Jan 1st 2000, how many times has SPY fallen 5% in a 7-day period? In other words, at time t, how many times has the percent return at time (t + 7 days) is -5% or more.
Note, I’m asking 7 calendar days, not 7 trading days.
In the results, include the data ranges of these drops and show the percent return. Also, format these results in a markdown table.
Here is a snapshot of the results from the model.
As a reminder, you can view all of the full results by clicking this link.
The question above starts on January 1, 2000, and the shown summary reports 150 qualifying seven-calendar-day windows. The original prose said 2020, although it also discussed 2008. Read back the linked query and rows to establish the actual end date and whether overlapping windows were counted separately before treating 150 as a confirmed count of downturn episodes.
Let’s dig a little deeper.
Analyzing how low the market can go
After a seven-calendar-day drop, inspect the exact measure requested for the next 180 and 365 days. Peak-to-trough drawdown, decline from the event’s starting price and return from the event’s ending price use different baselines. The transcript’s implemented query must settle that definition before the reported averages guide a rule.
How long can it really go?
Let’s find out.
The next question I asked sought to determine the drawdown within the next year. I asked the following:
From each of these start dates, what was the max drawdown within the next 180 days? What about the next 365 days?
Because the model outputted the full table, I asked it to summarize the data.
The shown summary reports an average subsequent one-year drawdown near 20%. Its sample needs the same January 1, 2000 start requested above, a confirmed end date and a stated overlap rule. Until the linked query and rows establish those details, keep this as the reported summary rather than a verified 2020-onward statistic.
Does this hint at the fact that you should sell when confronted with a little bit of volatility?
Not necessarily.
Analyzing what can happen if we continue to hold
From the above analysis, we know that markets can continue to crash during a large drawdown. But, that doesn’t mean we should sell.
To prove this, I’m going to ask the model the following.
From each of these end dates from the analysis, what was the percent return within the next 180 days? What about the next 365 days?
From this, we can see the average return half a year and a full year later. This data is too much for us to analyze, so we’ll similarly summarize it to get the key points.
What is the average 180 day return and the average 365 day return, and how does it compare to the 7 day percent drop?
From this, we can see something pretty interesting: the market tends to rebound after a steep drop.
Even though the maximum drawdown can be 2x to 4x the initial drop, we can see that the market tends to revert course.
The sample includes the 2008 financial crisis and COVID-19. Compare results with and without those episodes as a sensitivity check, keeping both sets. Removing them changes the question to behavior outside those crises; it does not by itself make the analysis more accurate.
For that comparison, use the same event definition and return horizon in both sets. Report their date ranges, counts and overlap rule alongside their averages.
Transforming these insights into trading algorithms
So far, I’ve shown you how to perform detailed financial analysis using an LLM tool. However, capturing the value of these tools is the most important aspect.
I’m going to show you how to translate this analysis into trading strategies.
Updating our experimental setup
To do this, I’m going to change the model to GPT-4o-mini. For tasks that don’t require reasoning, the GPT-4o-mini is more than capable of doing the task. This includes creating algorithmic trading strategies. It’s also much faster and cheaper.
I’ll also update the Quicktest settings so that the resulting backtests is more applicable for what I’m trying to test — a market-beating trading strategy at the height of a market downturn.
Creating our trading strategies
Next I will turn these ideas into candidate rules and compare their historical behavior with SPY.
Our trading strategies will capitalize on the fact that the markets tend to go up in the long-term, even in the face of large drawdowns. The rules I will choose are as follows:
- Buy 50% of our buying power in SPXL if we have less than $500 of SPXL positions
- Sell 20% of our portfolio value in SPXL if we haven’t sold in 10000 days and our positions are up 10%
- Sell 20% of our portfolio value in SPXL if the SPXL stock price is up 10% from when we last sold it
- Buy 40% of our buying power in SPXL if our SPXL positions are down 12% or more
This portfolio is MUCH more volatile than holding SPY and experiences drawdowns of up to 26% (a fact that I aligns with what we discovered earlier in this article).
However, if we held and kept holding, the leveraged strategy eventually goes up.
In the selected two-year backtest shown below, the saved rules outperformed the SPY comparison. That is the observed test outcome, rather than a promise that waiting two years will produce the same result.
And significantly outperforms the broader market.
The next screenshot shows another selected historical window. Compare its start and end dates, strategy return, SPY return and drawdown before generalizing to other bull markets.
Fair warning: this strategy has performed well in the past. This does not mean it’ll absolutely perform well in the future. Different market conditions result in different changes in stock price.
For example, if a nuclear war broke out or the AI robots take over, that could cause prolonged downturns in the market. This strategy counts on the fact that the market tends to go up in the long-term, and if this doesn’t happen, this strategy will do particularly poorly.
Now, if we do decide that this strategy was perfect for 2025, we could deploy it live. This could be paper-trading, a way to test a strategy live without risking actual money. Or, we could connect a brokerage, and see how this portfolio performs in the actual stock market.
Here’s how to deploy it live.
Step 1 (Optional): Create an Alpaca account and enable live-trading
This step is only necessary if we want to deploy our portfolio to the live market.
Within NexusTrade, there are two deployments: paper-trading and live-trading.
Paper-trading is like playing with monopoly money. It allows us to see how our portfolio performs without any risk. It’s our best form of testing, and gives a no-BS view of the performance of our strategy.
In contrast, live-trading is not a test. We are actually executing our trades in the stock market. When we’re ready to take this leap, we first have to do a little bit of setup.
First, we’ll navigate to the live-trading page.
Then we’ll click the Connect with Alpaca button and follow the instructions on the screen.
Once our Alpaca account is created and connected with our NexusTrade account, we can customize our portfolio with trading strategies.
Step 2): Add automated trading strategies to our portfolio
In this next step, we can add our strategies to our portfolio. We can do this for both paper-trading and live-trading depending on our specific goals. To do this, we’ll first click on the portfolio we created.
This opens a modal that allows us to update our strategies or add them to a portfolio. This includes creating a new portfolio, adding them to our live-trading Alpaca Portfolio, or adding them to a paper-trading portfolio.
After selecting a portfolio, we then click “Add to Portfolio” and our strategies will be ready to launch!
Now, it’s time to deploy our portfolio live to the market.
Step 4): Deploy our strategies to the market
Our final step is to actually deploy our portfolio, which is extremely easy. Regardless if we’re paper-trading or live-trading, we’ll navigate to our portfolio page, and click “Launch to Market”. This opens a modal and allows us to deploy our trading strategies.
Now, you’ve just launched a data-driven algorithmic trading strategy to the market. Congratulations!
Warnings and Disclaimers
This article would be incomplete without a warning: the past doesn’t necessarily predict the future.
The purpose of this article isn’t to convince you to go all in after the first biggest drop. It’s to show you how to perform data analysis to make more informed investing decisions.
The next screenshot is labelled Gemini Pro. Its displayed query uses LEAD(..., 7), which selects seven rows rather than a seven-calendar-day interval. In daily trading data those rows represent trading observations, so check the actual dates before accepting the result.
If we zoom in and examine the query, we’d notice something completely wrong with the query.
Instead of accurately counting 7 calendar days, it uses LEAD(…, 7). This explicitly gives us 7 trading days, which is not what we asked for.
Moreover, as I mentioned before, the past doesn’t necessarily predict the future. What happened during the 2008 Financial Crisis and the COVID-19 pandemic was unprecedented, and the resulting volatility of the market was too. If we were to enter into any sort of long multi-year depression, the strategies presented in this article would be particularly susceptible to drawdowns.
Nonetheless, this article still serves its purpose as a useful guide for people who want to get started with automated investing and financial research.
Concluding Thoughts
The purpose of this article is not to convince you to buy stocks during a drawdown or sell stocks because it can get much worse. It's to show you simple it is to perform stock analysis to empower you to make those decisions for yourself.
Thanks to the new OpenAI o1 model, investors, like you can perform accurate analysis with unprecedented speed, without having to write a single line of code.
You can translate the question into explicit rules and compare their historical outcomes with a benchmark. Inspect the query dates and saved trade events before accepting the result.
Overall, it gives you a glimpse of how to become a better investor by using a objective data. The time that O1 saved is phenomenal. reasoning is simply unmatched, and will transform how businesses utilize this new frontier of models.
And I couldn't be more excited to witness this.
If you made it this far, you MUST be interested in algorithmic trading or financial research. Thus, I ask you to please share this article on social media, and with a couple of your close friends. It helps spread my reach and directly supports me in accomplishing my mission to make Wall Street tools accessible to everyone, regardless of their financial or technical background.
And isn’t that a mission worth supporting?
Thank you for reading! If you’d like to keep up with my work or connect with me, here’s how.
LinkedIn | X (Twitter) | TikTok | Instagram | Newsletter
No comments yet.