Decision Making With Cognitive Financial Agents

AI Safety and Security in NASDAQ's Updated AI Policy.

Many people who try to implement an agent architecture for the first time think about tasks that can be solved through agent capabilities like reasoning, planning, and calculating (tool use). In reality, most implementors would be better advised to structure their agent architectures around which decisions need to be taken and then ensure that the agents have all the information they need to make that decision.

But what’s a “decision”?

Encyclopedia Britannica defines it like this:

1. a choice that you make about something after thinking about it: the result of deciding

2. formal: the ability to make choices quickly and confidently

3. the particular end of a legal or official argument: a legal or official judgment

4. : the act of deciding something

Wikipedia defines decision-making as “the cognitive process resulting in the selection of a belief or a course of action among several possible alternative options. It could be either rational or irrational.”

What all of these definitions have in common are the words “choice” and “end”.

Let’s begin.

Decision

Encyclopedia Autonomica is a blog about autonomous agents and the month of February 2025 was about applications in Finance. So this time again, my example will be to write a decision logic that can help a cognitive financial agent answer a question that gets posted on Reddit probably 100x daily

“Should I buy this stock?”.

While this sounds like an easy question to answer, especially for an AI Agent who has access to all the information in the world (it’s hyperbolic — yes I know), it is actually incredibly hard. Largely, because the contextual assumptions about investment horizon, location, risk appetite, and budget are not given. For example, if you have millions to invest you want to invest in a stock that likely has a limited up/down-side. On the other hand, if you only have little money, you might just want to chase your luck in stocks with high upward/downward risk, a profile usually seen in penny stocks. For the avoidance of doubt, I would not recommend this.

Like with all things in life, they work better if you have a clear concept about what you want to do.

Therefore, we will begin this study with a quick introduction of some characteristics of choice:

• P/E Ratio (Price-to-Earnings Ratio)
  Measures how much investors are willing to pay for each dollar of a company's earnings. This ratio helps assess whether a stock's price is justified based on its earnings.

• Debt-to-Equity Ratio (D/E Ratio)
  Compares a company’s total debt to its shareholders' equity, indicating how much of its operations are funded by debt versus owned capital. This metric is crucial in evaluating a company's long-term sustainability.

• Revenue Growth
  Revenue growth measures how a company's sales are increasing over time, typically compared year-over-year (YoY). A growing company is more likely to generate future profits and investor returns.

• Return on Equity (ROE)
  Measures how effectively a company uses shareholder equity to generate profits. This metric compares profitability across companies in the same industry.

• Insider Buying
  Insider buying occurs when company executives or board members purchase shares in their own company, which is often a sign of confidence in the stock’s future performance. Conversely, large-scale insider selling can sometimes indicate concerns about future growth or potential business challenges. While not a definitive signal, insider activity provides additional context for investment decisions.

So, how can we use this for decision-making then?

In its most simple form, we can define certain cutoffs. Contextual information that transforms the data point into information.

Here are some basic ideas of how this could look like:

For most of the characteristics, the approach is quite clear. But it already helps us make a decision.

How can we as humans use this to make a decision?

We can use our good old friend the decision tree. I already pruned it for better readability.

It’s a crappy image, but you’ll get the point. In its most basic form, the decision logic is an unweighted directional graph, where we stop at each red, don’t know how to proceed on the yellows, and move ahead on the greens. If you analyze all the 9000+ stocks on Nasdaq on a daily basis you might actually get a fine list of stocks you might want to invest in.

But wait, there is more. Because this technology already exists. This is called a screener. Screeners filter and rank stocks based on a selection of specific criteria, such as P/E ratio, revenue growth, or debt-to-equity ratio. Screeners are a handy tool to quickly narrow down a large pool of stocks.

This is an example of a screener on Yahoo Finance.

source

Why does this matter?

Context.

Again, it’s this word that mattered so much to me in the month of February. In our implementation we want our trading agent to make a decision. A choice. A conclusion. And end to the thought process.

Now if we have a given ticker, let’s say Nvidia (NVDA), we can use the same decision tree to evaluate it and reach a conclusion. In reality, we need to realize that while a decision tree is fun and easy, it doesn’t scale well with the complexities and dynamism of the investment decision domain. Here current-tech implementations are decision engines — that I had covered in a previous post.

Now, where do agents come in?

Cognitive Financial Agents

Financial decision-making isn’t only a quantitative discipline. Even in our example above we evaluated a qualitative measure if the Executive team is buying or not. But stock markets have more dimensions of complexity, here we are making decisions in an environment where information is incomplete, unstructured, and constantly changing — and sometimes adversarial. The complexity of our almost always-on modern financial markets goes far beyond what a person with a decision tree could handle.

The reality is that all Financial institutions face an overwhelming volume of dynamic data from earnings calls, regulatory filings, and market signals. And especially for the smaller players, this is extremely hard to manage. On top of that, AI policies like the newly published one from NASDAQ, create operational inefficiencies that have unearthed bottlenecks in data analysis, compliance monitoring, and risk assessment—leading to errors, increased costs, and missed opportunities.

This is where Cognitive Financial Agents (CFAs) come in. The way I think about them, CFAs are autonomous agent systems designed to enhance decision-making efficiency, improve risk management accuracy, and ensure regulatory compliance—all while reducing operational costs. Rather than relying on rigid, rule-based systems, CFAs operate through an interconnected Economic Consciousness Network, dynamically adapting to new financial realities.

Core Functionalities of How CFAs Will Work

• Advanced Reasoning & Learning – CFA’s analyze, synthesize, and adapt financial strategies in real-time.

• Autonomous Decision-Making – Search, extract, and analyze financial data to optimize investments.

• Risk Management & Compliance – Continuously monitor financial portfolios, adjust exposures, and ensure regulatory adherence.

• Goal-Oriented Planning – Negotiate and strategically plan portfolios and adapt dynamically to new economic trends.

• Governance & Oversight – Built-in safety mechanisms ensure AI decisions align with business and regulatory standards.

In my mind, by leveraging these capabilities, CFAs will transform fragmented financial information into structured, actionable decisions—empowering financial institutions to operate with greater precision, speed, and intelligence.

If the governance function works out. And, tbh, that is the component where I have the most concerns right now. I buy a lot of data from Nasdaq, and recently they updated their AI Policy.

NASDAQ’s AI Policy

Dissemination of Nasdaq Information, including when used in large language models (LLM), retrieval-augmented generation (RAG), or other AI functionality, must be controlled through a technical entitlement system that can be interrogated for audit and compliance purposes.

AI Safety

As you can see, the policy emphasizes responsible use and governance of financial data in AI systems, outlining strict obligations for AI model training, third-party tool usage, operational controls, and licensing. More importantly, AI models trained on Nasdaq data must adhere to licensing agreements, preventing open-source access or use in high-risk AI environments. This means that open-source AI where the code and training data are publicly accessible and modifiable is not permitted. Because allowing Nasdaq data in such models could lead to uncontrolled distribution and misuse. Fair enough, I guess. Secondly, some applications are deemed high-risk/dangerous or unethical under laws like the EU AI Act, such as AI systems used for illegal surveillance, social scoring, or generating misleading financial advice.

Then, third-party AI tools, like my CFA, would likely interact with Nasdaq data that would require appropriate permissions. More importantly, producing derivative works or aggregated data cannot be distributed without a license. This would be especially important if I would relaunch DeepCQ as a product.

AI Security

Operational controls mandate strict access monitoring, audit trails, and safeguards against unauthorized redistribution. Organizations using Nasdaq data in AI systems must implement entitlement tracking and prevent data reconstruction or unauthorized replication. Any misuse or unauthorized distribution must be reported, and Nasdaq reserves the right to revoke access and require data purging upon contract termination. These governance measures align with evolving regulatory standards, including the EU AI Act, ensuring transparency, accountability, and data integrity in AI-driven financial systems.

In closing

I see great potential in building that agent, but rest assured the existence of such tools in a highly regulated environment like investing can lead to governance issues. But tbh, in markets where high-frequency trading, leveraged options, and hedge funds exist, governance should be taken seriously, but it can’t be a showstopper.

Superbill is waiting. I’m excited.

Hope you like the post. Please share and subscribe.