Agentic AI Financial Analysis with Autogen: What is Better, Bitcoin or Tesla?

Author(s): Igor Novikov

Originally published on Towards AI.

Have you ever wondered which is a better investment — NVidia or Tesla? But what if you are too lazy to do the analysis yourself or don’t even know how to begin?

A serious problem, I know. How awesome would it be if we could make ChatGPT do that but if you have already tried — you know it is not very good at it. You can ask it to make a plan but it’s hard to make it follow it through multiple steps. And it is not very good accounting so analyzing financial data is not its thing yet. Can something be done about it or is there no hope and we are doomed to spend our time and hard-earned money in Hamster Combat?

And the answer — there is a solution, albeit not as simple as ChatGPT but as capable of overcoming many of the ChatGPT or even more complex RAG and LLM-based solutions.

I’m talking, of course, about agentic AI. The idea is that we have a group of moderately intelligent but highly functional agents that through constructive dialog and argument will figure out what is what. As they say, discussio mater veritas est.

Talking seriously though, this approach is somewhat similar to boosting in ML. We don’t need agents to be super-intelligent, instead, we need a lot of agents that are specialized and have appropriate tools for tasks they need to accomplish.

For example, for the task above we need the following:

• A mediator will mediate the conversation
• A financial analyst who can guide how to assess different assets’ performance, and can do so based on certain policies
• A coder that will program the algorithm based on analyst description.
• A critic will evaluate the answer and provide feedback. This is an important component in reducing hallucinations

Those are agents of the system. Let’s decompose that into LLMs architecture components:

• Mediator — this is a LLMs agents group manager. It needs to know what is the task, who are the participants, and how the conversation is structured. Could be LLM based.
• Financial analyst — you likely want it to you YOUR investment policies and documents. If you have a lot — that’s a RAG system, if not — we can just feed that to the LLM as part of the context
• A coder — that’s an LLM code generator and execution module. It should support retries and the ability to fix its own errors. Ideally, it should run in an isolated environment like a docker container
• A critic — let’s assume we will use LLM as a judge approach, so it will be an LLM agen with defined criteria

Looks difficult already… but it’s not. Here into play come frameworks. It seems all popular frameworks seem to converge on the idea of agents nowadays, with both LLamaIndex and Langchain having them. I’m going to use a Microsoft open-source framework called Autogen.

Agentic approach

So what is an agentic framework? It is simply a combination of the following:

LLM + Code Interpreter + Plugins

AutoGen was created to simplify building AI agent systems that can run with or without human input. What can it do?

1. Call functions
2. Read documents
3. Have short and long-term memory
4. Create and organize different agents’ conversations on certain topics
5. Learn things as it goes
6. Use provided tools
7. Organize RAG systems
8. Generate and execute code
9. Optimize usage of different agents through training
10. Integrate with existing low-code and no-code platforms

That’s all we need so let’s build a simple system, that can answer once and for all, what is better — Bitcoin or Dogecoin?

We are going to need an OpenAI API key, as we are going to use gpt4o as a model. You can use any model with Autogen, including self-hosted but for simplicity, I’ll go this path.

Once you have the key we do the following configuration:

config_list = [
 {
 'model': 'gpt-4o',
 'api_key': 'DER PAROL'
 }
]


llm_config={
 "timeout": 600,
 "config_list": config_list,
 "temperature": 0
}

Now we create our agents.

Our financial consultant:

assistant = AssistantAgent(
 "consultant",
 llm_config=llm_config
)

Our coder:

code_interpreter = UserProxyAgent(
 "coder",
 human_input_mode="NEVER",
 code_execution_config={
 "work_dir": "coding",
 "use_docker": False,
 },
 default_auto_reply=""
)

And our critic:

critic = AssistantAgent(
 "critic",
 llm_config=llm_config,
 system_message = "You are a critic, analyze the data, the solution proposed and made conclusions and indentify problems rated from 1 to 10 by severity. If any problem has severity more than 8 - ask to fix it"
)

They are going to use gpt4o for general thinking and Python interpreter for coding.

Now let’s organize them into a thinking group, there is a special class for that called SocietyOfMindAgent:

def init_group(assistants):
 groupchat = GroupChat(
 agents=assistants,
 messages=[],
 speaker_selection_method="round_robin", # With two agents, this is equivalent to a 1:1 conversation.
 allow_repeat_speaker=True,
 max_round=8,
 )


 manager = GroupChatManager(
 groupchat=groupchat,
 llm_config=llm_config,
 )


 group = SocietyOfMindAgent(
 "society_of_mind",
 chat_manager=manager,
 llm_config=llm_config,
 )


 return group


society_of_mind_agent = init_group([assistant, code_interpreter, critic])

We also need an object repesenting us, a user:

user_proxy = UserProxyAgent(
 name="Human_admin",
 human_input_mode="NEVER",
 code_execution_config=False,
 max_consecutive_auto_reply=0
)

That’s basically it. We have a group chat so let’s test it on a question we are all interested in:

task = "Which stock is a better investment for a conservative investor that is trying to maximize long term gains - NVDA or TSLA based on performance over last 15 years"
user_proxy.initiate_chat(society_of_mind_agent, message=task)

The output looks a bit like magic because it is a conversation between several agents. It’s a long one so the short summary is:

1. It was able to come up with a method to compare both stocks.
2. Write a code to fetch needed financial data from the internet
3. Execute the code and get the data
4. Get in error — fix it
5. Execute the code again, successfully
6. Get the aggregated financials and analyze them
7. Make conclusions

Here are they btw:

Based on the calculated metrics for NVDA and TSLA over the last 10 years, we can make the following observations:

### Analysis

1. **Total Return**:

– NVDA: 26,767.72%

– TSLA: 3,197.86%

– NVDA has significantly outperformed TSLA in terms of total return over the last decade.

2. **Annualized Return**:

– NVDA: 74.95%

– TSLA: 41.85%

– NVDA also has a higher annualized return compared to TSLA, indicating a better average yearly performance.

3. **Volatility**:

– NVDA: 48.58%

– TSLA: 57.00%

– NVDA has lower volatility compared to TSLA, suggesting it is less risky in terms of price fluctuations.

### Conclusion

For a conservative investor looking to maximize long-term gains, NVDA appears to be the better choice based on the historical data. It offers a higher annualized return and lower volatility compared to TSLA, aligning more closely with the goals of a conservative investment strategy.

### Identified Problems

1. **Data Limitation** (Severity: 3): The analysis is based solely on historical data, which may not predict future performance.

2. **Market Conditions** (Severity: 4): Changes in market conditions or company-specific factors could impact future performance.

3. **Volatility Consideration** (Severity: 5): While NVDA has lower volatility than TSLA, it is still relatively high, which may not suit all conservative investors.

Based on the performance over the last 10 years, NVDA (NVIDIA Corporation) appears to be a better investment for a conservative investor aiming to maximize long-term gains compared to TSLA (Tesla, Inc.). NVDA has demonstrated a significantly higher total return of 26,767.72% and an annualized return of 74.95%, compared to TSLA’s total return of 3,197.86% and annualized return of 41.85%. Additionally, NVDA has lower volatility at 48.58% compared to TSLA’s 57.00%, indicating less risk in terms of price fluctuations. These factors suggest that NVDA aligns more closely with the goals of a conservative investment strategy. However, it’s important to consider that past performance does not guarantee future results, and market conditions can change.

Nice and I always knew that NVDA is a less speculative choice, but I’m not good at reading reports, I’d like it to draw a chart for me. Can it do it?

task = "I would like to compare NVDA and TSLA performance for the last 10 years. Suggest a method to do the analysis and perform it. Plot results to a chart and save it to a file chart.png"
user_proxy.initiate_chat(society_of_mind_agent, message=task)


Image(filename="coding/chart.png")

Sure it can. That is short of amazing!

Even better. But I being a very intelligent investor read a book recently, which unsurprisingly, is called Intelligent Investor by Benjamin Graham and it is now my Bible in investing. So I want my system to base all its decisions on what Graham recommends. Can we do that?

Yes we can, now we need to build a RAG system, so we can store our new investment policies and other documents in a knowledgebase. We are going to use Qdrant vector database as storage:

sentence_transformer_ef = SentenceTransformer("all-distilroberta-v1").encode
client = QdrantClient(":memory:")


ragproxyagent = RetrieveUserProxyAgent(
 name="ragproxyagent",
 human_input_mode="NEVER",
 max_consecutive_auto_reply=10,
 retrieve_config={
 "task": "code",
 "docs_path": [
 "https://youexec.com/book-summaries/the-intelligent-investor-by-benjamin-graham"
 ], 
 "chunk_token_size": 2000,
 "model": config_list[0]["model"],
 "db_config": {"client": client},
 "vector_db": "qdrant", 
 "get_or_create": True, 
 "overwrite": True, 
 "embedding_function": sentence_transformer_ef, 
 },
 code_execution_config=False,
)

And the answer is:

To determine which stock, NVDA (NVIDIA) or TSLA (Tesla), is a better investment for an intelligent investor as defined by Benjamin Graham, we need to analyze both companies based on Graham’s principles of value investing. This involves evaluating their financial stability, earnings growth, dividend record, and valuation metrics such as the price-to-earnings (P/E) ratio and price-to-book (P/B) ratio.

### Analysis:

1. **Financial Stability**:

– **NVIDIA**: Has a total debt of approximately $11.06 billion and cash and cash equivalents of $7.28 billion. The net debt is $2.43 billion, indicating a relatively stable financial position with a manageable level of debt.

– **Tesla**: Has a total debt of approximately $9.57 billion and cash and cash equivalents of $16.39 billion. Tesla appears to have a stronger cash position relative to its debt, suggesting good financial stability.

2. **Earnings Stability and Growth**:

– Both companies have shown significant earnings growth in recent years, with NVIDIA benefiting from the demand for GPUs and Tesla from the electric vehicle market.

3. **Dividend Record**:

– Neither NVIDIA nor Tesla is known for paying significant dividends, as both companies have historically reinvested earnings into growth and expansion.

4. **Valuation Metrics**:

– Both companies are often considered growth stocks and may have high valuation multiples, which could be a concern for value investors.

5. **Margin of Safety**:

– Given the high growth expectations priced into both stocks, a value investor might find it difficult to justify a significant margin of safety.

### Conclusion:

Based on Benjamin Graham’s principles, neither NVIDIA nor Tesla perfectly fits the criteria for a value investment due to their high growth valuations and lack of dividends. However, Tesla’s stronger cash position relative to its debt might make it a slightly more stable choice in terms of financial health. Nonetheless, both stocks may not align well with the conservative, value-focused approach advocated by Graham.

Well said.

And here is for Bitcoin:

task = "I would like to compare BITCOIN and DEGECOIN performance for the last 5 years. Suggest a method to do the analysis and perform it. Plot results to a chart and save it to a file chart.png"
user_proxy.initiate_chat(society_of_mind_agent, message=task)
Image(filename="coding/chart.png")

Augtogen provides much more functionality than described here. This is a very basic example really but shows that you can do a lot with little code.

Additionally, there is AutogenStudio, which can be self-hosted and allows you to do all that in a simple UI with little coding.

Conclusions

Agentic systems are quite more powerful than ChatGPT alone or even advanced RAG systems. Unfortunately, there are some problems. The output is unstable, and you have to run the same question several times before it gets you a decent answer. With Autogen specifically, sometimes agents are stuck in a loop or hallucinate. And Microsoft being quite itself — the framework looks nice but is unstable and has bugs, making it hard to use for production systems. It is quite nice for building rapid prototypes though.

One thing is clear though, that this is the future of intelligent decision-making systems as it makes multistep logic quite more natural and controllable than working with ChatGPT or similar apps.

Here is a notebook with the code from the article.

Nothing in this article is financial advice, all examples are for demonstration purposes only!

Have fun! If you have any questions — put them in a comment.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI