Introducing Financial Recommendation

What and Why is it Interesting?

Dr. Richard McCreadie

SIGIR - 13/07/25

(Open Market) Investing

​

​

Problem and Challenges

Finding the Right Investments is Hard

• Complexity: Finance is complicated, and many people are not sufficiently educated to understand the consequences of their decisions (…and the large volumes of financial jargon does not help!)
• Time: Investing successfully takes time and effort to research and understand the target markets, most people don’t have the time for this
• Risk: There are a wide range of investment risks, and some of those are difficult to effectively quantify
• Choice: The range of possible investments is so large that choice paralysis is a barrier
• Advice: Its not clear to a new investor where the should go to get advice, and who they can trust

Investment Recommendation

• Given these problems, we cannot expect an average member of the public to become a savvy investor on their own
• They need a financial advisor to analyze their position and recommend assets to invest in personalized to them
• Manage investment risk to the customer by identifying profitable assets that meet their risk profile

​

• … but expert financial advisor time is limited and expensive
• Also, may only specialize in particular asset types or markets

Robo-Advisors

• Robo-advisors are computer programs that can analyse customer and market data and provide financial advice
• First devised around 15 years ago
• Increasingly common, with companies like Betterment, Wealthfront, and Personal Capital managing billions in assets
• Most are not very sophisticated, i.e. based on rules or basic pattern analysis

​

5

Recommender Systems are a core component of these platforms!

Financial Recommendation

What is Recommendation?

• Recommender systems are fundamentally concerned with the process of ranking items for users based on one or more suitability criteria

Example:

Movie Recommendation

User

(Profile)

Scoring Function: .

Movie Genre Similarity to User Profile

• Items: Movies (Title, Description, Genre, Actors…)
• User: A representation of what the customer likes to watch (Watched Movies, Category Preferences, Skipped Recommendations)
• Suitability Criteria: Genre Similarity (Usually an aggregate of multiple signals)

Financial Recommendation

• In financial recommendation we are ranking financial assets for potential investors based on how suitable they are for that investor

Financial Recommendation

Investor

(Profile)

Scoring Function: .

Profitability

• Items: Financial Assets (Stocks, Bonds, Options, Funds, Derivatives)
• User: A representation of what the customer’s investment needs and preferences (Past Investments, Risk Appetite, Expertise, Horizon)
• Suitability Criteria: Profitability (…usually, we will come back to this…)

What Makes Finance Different?�(and interesting!)

• Conceptually, finance looks like any other recommendation task, but it does not act like a normal recommendation task in practice (and classical recommendation strategies do not work)

​

• Why?

Investment Suitability is Difficult to Measure

User behaviour is not a strong signal

Temporal dynamics have a strong impact

…and there are not good surrogates

…meaning that most user data is difficult to use

…so we can’t just ignore time

Investment Suitability

Its complicated…

• What do we even mean by investment suitability, lets start with a legal definition
• [A financial advisor must] “have a reasonable basis to believe a recommended transaction or investment strategy involving a security or securities is suitable for the customer. This is based on the information obtained through reasonable diligence of the firm or associated person to ascertain the customer’s investment profile.”

​

​

​

​

​

• That’s a bit more complicated than ‘I liked it’ or ‘I bought the thing’, and raises important questions, such as:
• How do we collect the basic information we need about an investor?
• How do these different criteria impact the scoring of an asset?
• Does it even make sense to consider items independent of a portfolio/basket of assets?
• Where can I get a crystal ball to see into the future with?

​

Investment Suitability is Difficult to Measure

“…includes, but is not limited to, the customer’s age, other investments, financial situation and needs, tax status, investment objectives, investment experience, investment time horizon, liquidity needs [and] risk tolerance,” among other information.”

FINRA: Rule 2111

• Because trying to measure true investment suitability is so difficult, most works opt for a simpler alternative… profitability
• Pricing data for financial assets are publicly available, so it is possible to simulate investment options over past time frames
• …so many works recommend assets that they predict will provide a good return on investment

Investment Suitability is Difficult to Measure

TODO: Citation

Price

Next-generation personalized investment recommendations. Big Data and Artificial Intelligence in Digital Finance, 2022

• If you come from a traditional recommender systems background you might think there is another alternative, could we use actual customer investments as a surrogate signal?
• If a user invested in an asset then surely they are happy with it?
• This would allow us to apply traditional recommender approaches like matrix factorization, since investments could act as ‘clicks’ or ‘likes’ by the user
• We could also use traditional recommender/ranking metrics for evaluation

Investment Suitability is Difficult to Measure

TODO: Citation

• While this is a nice theory, it runs into two practical problems
1. There are very few datasets that provide financial transactions that can be used
2. Transaction-based metrics do not correlate well with profitability
• … and it is difficult to argue we should prioritise models that are less profitable

Profitability

Transaction Suitability

On transaction-based metrics as a proxy for profitability of financial asset recommendations, FinRec 2022

Investors

And how it is difficult to understand them…

• In most recommendation settings we can rely on users to provide signals that our system is succeeding or failing (that we can learn from)
• Active: Providing star ratings on movies or buying products and writing product reviews
• Passive: Scrolling past items shown, not selecting an item when on its product page

​

• But this is less true in finance, as the users lack intuition regarding assets

User behaviour is not a strong signal

Just from this movie poster you likely have some idea of whether you would like to watch this movie…

…and if you watched it you would be able to rate it accurately afterwards

Most people would not be able to interpret what all of this information means…

… and this does not change after investment, since they may not be able to determine whether it is performing well for months or years later

Are Generative AI Agents Effective Personalized Financial Advisors? SIGIR 2025 (Talk Tomorrow, PALLADIO, ~12:15, Domain-Specific Applications 1 Session!)

• There are also challenges in interpreting the actions of investors, an investment could occur for a variety of reasons
• The user is an expert and is confident that this item is a good long-term investment
• The user is a day-trader and is re-balancing their portfolio based on the news of the day
• The user is a bot and is reacting to micro price movements and statistical signals
• The user is a bank customer and has been sold an investment by the bank
• A major event might have occurred that required re-evaluation of the market
• … and so on

​

• The issue here is that we typically have little information on why an asset was purchased or sold, and that information is important if we want to learn investment patterns

​

User behaviour is not a strong signal

TODO: Citation

Beyond Profit: Evaluating Financial Recommenders with Real-World Transactions, FinIR Workshop 2025. (Keynote at 8:40am in PALLADIO B Thursday!)

Time

It makes everything more challenging…

• Most traditional recommendation works ignore time as an element, they just assume that we will periodically re-train model with recent data
• This is based on the assumption that user preferences only change slowly

​

• However, the finance domain is far more volatile and is impacted by real world events
• This means that we need fall-back options if our models stop working

Temporal dynamics have a strong impact

US Stock Market (last 18 months)

• If you recall the legal definition I mentioned earlier, one of the features that was mentioned was the ‘investment time horizon’
• This is how long the user wants to invest their money for (weeks, months, years, decades)

​

• The most common way to lose money during investment is to be forced to sell your position when an asset is performing poorly
• Let’s imagine I bought some Nvidia stock in Oct 24
• For the rest of the year it looks like a strong investment
• But if I need to liquidate my investments to buy a car in

Apr 25, then I would have lost money

​

​

Temporal dynamics have a strong impact

Nvidia Stock Price

• Financial recommender systems need to factor in the investment time horizon, as well as consider the risk that the user needs to exit their position early

• In traditional recommender settings, an asset is largely static – a movie does not suddenly change after release

​

• But relevant aspects of a financial asset might fundamentally change
• Stocks represent companies, and those companies may change their business model
• The value of government bonds reflect confidence in that country
• Currency value represents the demand for that currency on public markets

​

​

Temporal dynamics have a strong impact

Netflix Stock Price

DVD Rental Company

Streaming

Transition

Streaming First

We need You!

… to help solve these problems

Outlook

• Financial Recommendation is one of the most impactful and challenging recommendation domains, but has seen relatively little exploration so far

​

​

​

​

• What makes Financial Recommendation challenging is that we need to consider a wider range of competing factors than traditional recommender systems
• Investment Suitability is difficult to measure and is multifaceted
• Investors have complex competing needs and are heterogeneous
• Understanding the impact of temporal dynamics is a critical factor in succeeding at recommendation

​

​

​

​

This makes Financial Recommendation a rich domain for research

Questions?

Q&A