Anupama Venugopalan

Case study · Explainable AI

Explainable Portfolio Playground

Can explanations change how people judge an AI suggestion?

Interactive prototypeExplainable AIHuman-AI decision making

This project explores whether explaining an AI recommendation changes how people understand it and how much they trust it. I was interested in a question that goes beyond whether an AI system can produce a recommendation: can the person using it understand the reasoning, question the suggestion and make their own decision?

To explore this, I built an interactive portfolio experiment in Streamlit. Participants experience either a black-box recommendation or an explainable version of the same process. The app collects their investment preferences, suggests a five-stock portfolio, presents possible five-year outcomes and asks them to evaluate how much they trust the recommendation.

Welcome screen explaining the steps in the black-box and explainable AI experiment.
The welcome screen introduces the participant journey and frames the prototype as an experiment about AI explanations and trust.

Try the live Streamlit prototype

The experience

The app suggests one stock from each of five sectors. For every sector, the participant can accept the AI suggestion or select the alternative stock. The projected portfolio results then reflect those choices. At the end, the participant rates how much they trust the portfolio and how likely they would be to act on it.

Investor profile and black-box versus explainable mode selection in the Streamlit prototype.
The participant sets a profile and chooses whether the recommendation will show explanations.

What the black-box view leaves out

In black-box mode, the user sees the selected portfolio and its overall scenarios without detailed reasons for each stock. Trust and willingness to follow the suggestion are then rated on separate scales. This makes the absence of explanation visible as part of the experience.

Five-stock portfolio in black-box mode with no reasons shown, followed by trust rating controls.
The same portfolio can be shown without reasons. The trust controls capture the participant's response.

What the explainable view adds

The explainable view shows plain-language reasons alongside three displayed factors: 12-month return, 60-day volatility and average correlation. The user can compare why a stock appears in the portfolio, while retaining the ability to choose another option.

Explainable portfolio view showing a reason for each selected stock and numerical factors.
The explanation table pairs each selection with factors and a short reason that a user can examine.
Charts comparing selected stocks by return, volatility and average correlation.
The visual comparison makes the factors behind the recommendation easier to inspect across the five selected stocks.

Making uncertainty visible

The five-year chart uses Monte Carlo simulations to show pessimistic, median and optimistic paths. These are illustrative model scenarios rather than reliable predictions of future returns. In the case study, they demonstrate how a decision interface can communicate a range of possible outcomes.

Five-year Monte Carlo portfolio chart and table of illustrative year-five outcomes.
The interface places a range of possible outcomes next to the selected stocks instead of presenting one certain number.

Trust comparison

Explainable AI received higher trust ratings.

Average trust in the AI-assisted recommendation, rated from 0 to 10

Black-box AI
3.9
Explainable AI
7.9
+4.0 pointsdifference in average trust rating

What I learned

An explanation is useful only if it helps someone question the recommendation, not merely accept it. The submitted ratings show a clear difference between the two experiences: average trust was 7.9 for explainable AI and 3.9 for the black-box version. The result supports further investigation into how explanation design influences trust and decision quality.