I recently had the chance to speak with MBA students at the University of Economics about the vital role of business experts in data and machine learning applications. Rather than focusing on technical details, I emphasized how their industry knowledge and strategic thinking can drive successful ML projects. Through a case study on a risk manager using ML for a scorecard project, I illustrated how business professionals can leverage data-driven insights without needing deep technical expertise. In this blog, I’m sharing my slides and key insights to help others understand how to apply ML effectively in a business context.
Q&A:
#1 —————————————————
Student asked: “I’m working in a fintech company, offering stock trading, what is the variables of the ML model?”
Tri answered:
First, the question should be reframed. Instead of asking, “What are the variables of the model for a fintech company?”, you should ask:
“I want to protect and grow the business. How can I apply Machine Learning?”
Step 1: Understand Your Business Portfolio & Revenue Sources
Before jumping into ML, you need to analyze your business fundamentals:
. What are the sources of income? (e.g., trading fees, spreads, subscriptions, margin lending)
. Who are your customers? (e.g., retail traders, institutional investors, high-frequency traders)
. How do they behave? (e.g., trade volume, retention rate, churn rate, deposit patterns)
Step 2: Analyze Data Across Multiple Dimensions
Look at historical data from different perspectives:
. By stock: Which stocks drive the most trading activity?
. By customer segment: Are high-frequency traders more profitable than long-term investors?
. By demographic: Are younger users trading more actively than older ones?
. By trend: Is there a slowdown in a particular segment? Are certain products underperforming?
Step 3: Identify Key Business Challenges & Make a Case for ML
Once you detect issues like a decline in trading volume, an increase in churn, or low engagement in certain segments, you can start building a business case:
. Why do you need ML?
. What problem will it solve?
. How will it increase revenue or reduce risk?
Step 4: Discuss with Management & Defend Your Idea
. Present your findings and explain how ML can be a strategic tool:
. For retention: ML models can predict which users are likely to stop trading and suggest personalized offers.
. For growth: AI-driven recommendations can increase trading activity.
. For risk management: Fraud detection models can prevent financial losses.
Step 5: Dive into ML Research (or Ask ChatGPT!)
At this point, you can explore relevant ML models:
Read industry papers, case studies, or competitor strategies.
Use AI (like ChatGPT) to get insights on which ML models suit your business needs.
Onboard vendors (3 quotations from 3 vendors / of your in-house modelling team) then you can learn of their proposals.
Takeaway: The right approach is business-first, data-driven, and ML-enabled—not just focusing on model variables but understanding the real business impact of ML. 
Now, let’s check with ChatGPT further regarding the model:
Prompt:
I’m working in a fintech company offering stock trading service to customer with flexible trading amount and volume,
1. what is the main source of revenue for my business model?
2. suggest me how to apply Machine Learning to grow the business, retain customer with the detail case study.
ChatGPT:
1. Main Sources of Revenue for Your Fintech Stock Trading Business
Your company likely generates revenue from several key streams:
- Trading Commissions & Fees: Charging customers a fee per trade (fixed or percentage-based). Some platforms offer commission-free trading and monetize in other ways.
- Spread Revenue: Making money on the bid-ask spread, particularly in market-making or fractional share trading.
- Subscription or Premium Services: Offering advanced tools, analytics, research, or priority support for a monthly/annual fee.
- Margin Lending & Interest Income: Earning interest on margin loans when customers trade with borrowed money.
- Payment for Order Flow (PFOF): Earning fees from market makers or exchanges for routing customer orders.
- Asset Management Fees: If offering robo-advisory or managed portfolios, charging a percentage of assets under management (AUM).
- Stock Lending Revenue: Lending out customers’ stocks to short sellers and earning a lending fee.
- Partnerships & Advertisements: Monetizing through partnerships with financial service providers, ETFs, or other fintech firms.
2. Applying Machine Learning to Grow the Business & Retain Customers
Case Study: Personalized Trading & Retention Strategy
Problem Statement
Your platform has a large customer base, but retaining users and keeping them engaged is a challenge. Many users sign up but do not actively trade, while others stop after a few transactions.
Machine Learning Application
Step 1: Customer Segmentation & Churn Prediction
- Model: Use clustering (e.g., K-Means, DBSCAN) or supervised learning (e.g., Random Forest, XGBoost) to segment traders into different categories—active traders, casual traders, and dormant users.
- Features: Trading frequency, portfolio size, deposit patterns, risk tolerance, past churn behavior.
- Outcome: Identify users likely to stop trading in the next 30 days and target them with retention strategies.
Step 2: Personalized Investment Recommendations
- Model: Use collaborative filtering (like in recommendation systems) to suggest stocks or ETFs based on users with similar profiles.
- Features: Past trades, industry preferences, news sentiment analysis, technical indicators.
- Outcome: Increase engagement by recommending stocks aligned with user preferences.
Step 3: AI-Powered Trading Assistant
- Model: NLP models (e.g., ChatGPT-like models) to provide trade insights, explain market trends, and answer user queries.
- Features: Real-time market data, user queries, trading patterns.
- Outcome: Boost engagement by providing intelligent market insights via chatbots.
Step 4: Fraud & Anomaly Detection
- Model: Anomaly detection algorithms (Isolation Forest, Autoencoders) to detect unusual trading behavior, insider trading, or bot activities.
- Outcome: Improve platform security and compliance, maintaining customer trust.
Step 5: Predictive Portfolio Management
- Model: Reinforcement learning to optimize portfolio allocation based on market trends and user risk profiles.
- Outcome: Help users maximize returns and keep them engaged with an intelligent portfolio tool.
Expected Impact
- Higher Retention: Personalized recommendations and alerts keep users engaged.
- Increased Trading Volume: AI-powered insights encourage users to trade more.
- Enhanced Customer Experience: AI chatbots and assistants improve service quality.
- Lower Churn Rate: Early identification of at-risk customers allows for proactive retention strategies.
Would you like me to refine any of these ideas further, such as building a proof of concept or detailing an implementation plan?
Prompt:
How these models, at the end, contribute to the revenue?
ChatGPT:
Great question! Let’s directly connect each machine learning (ML) model to revenue generation.
1. Customer Segmentation & Churn Prediction → Higher Retention → More Trading Fees
- How it Works: Identifies users likely to stop trading and targets them with personalized incentives (e.g., discounts, exclusive insights, free trades).
- Revenue Impact: Reduces churn, keeping customers actively trading and generating more commissions, spreads, and order flow revenue over time.
2. Personalized Investment Recommendations → Increased Trading Volume → More Transaction Revenue
- How it Works: Recommends stocks, ETFs, or trading strategies based on user preferences and market trends.
- Revenue Impact: Encourages users to trade more often, increasing trading commissions, spreads, and payment for order flow (PFOF) revenue.
3. AI-Powered Trading Assistant → Higher Engagement & Upselling → Subscription & Margin Lending Growth
- How it Works: Provides real-time market insights, news, and trading tips via chatbot or voice assistant.
- Revenue Impact: Keeps users engaged, leading them to subscribe to premium services (e.g., advanced research, AI-powered alerts) and use margin lending, which generates interest revenue.
4. Fraud & Anomaly Detection → Risk Mitigation → Protects Revenue & Trust
- How it Works: Detects insider trading, bot activity, and fraudulent transactions, reducing compliance risks.
- Revenue Impact: Prevents financial losses and regulatory fines, ensuring smooth business operations and maintaining customer trust, which indirectly increases revenue.
5. Predictive Portfolio Management → Better Performance → Asset Management & Advisory Revenue
- How it Works: Uses reinforcement learning to optimize portfolios, helping users make better investment decisions.
- Revenue Impact: Increases AUM (Assets Under Management), generating higher advisory fees and premium subscriptions for automated portfolio services.
Summary: How ML Directly Contributes to Revenue
| Machine Learning Model | Revenue Impact |
|---|---|
| Churn Prediction | More retained traders = More trading fees |
| Investment Recommendations | Higher trading volume = More commissions & spreads |
| AI Trading Assistant | More engagement = More premium subscriptions & margin lending |
| Fraud Detection | Prevents financial loss = Maintains customer trust & compliance |
| Portfolio Optimization | Higher AUM = More advisory & management fees |
By implementing these ML strategies, you can increase customer lifetime value (CLV), boost trading activity, and unlock new revenue streams.
Would you like a deep dive into a specific model or an implementation plan?
#2 —————————————————
Student Asked: Should I read a paper, and use that as a base to develop my model at my business, e.g. use the same set of variable, hypothesis, etc.?
Tri Answered: Not really, you should start from business insight, paper can be used as reference source, but not the golden source where you can get the variable.
Check with ChatGPT:
1. Define the Business Problem Clearly
Before choosing independent variables, you must be clear on:
- What business outcome (dependent variable) you want to predict or optimize.
- Example: Customer default probability, trading volume prediction, or fraud detection.
- How this outcome impacts the business.
- Example: If predicting customer churn, reducing it can increase retention revenue.
- What actions the business can take based on the model’s predictions.
- Example: If predicting trading volume, can you design better promotions to increase activity?
2. Brainstorm Potential Drivers of the Outcome
Since you are working in fintech with stock trading services, consider what factors logically influence the outcome you’re predicting. Use domain knowledge and hypothesis-driven thinking to brainstorm potential independent variables.
Example: Predicting Trading Volume
Possible independent variables:
- User behavior: Number of past trades, frequency of logins, deposit amounts.
- Market factors: Stock index movements, volatility (VIX), interest rates.
- User demographics: Experience level, portfolio size, risk tolerance.
- App engagement: Time spent on the platform, clickstream data.
Start broad—then refine based on what’s available.
3. Check Data Availability
Now, validate which of these variables you actually have access to. Sometimes, the most impactful variables are not collected or not easily available.
- Internal data: Do you track user behavior, trade history, deposits, and withdrawals?
- External data: Can you access market trends, economic indicators, social sentiment?
- Feature engineering: Can you derive useful insights from raw data? (e.g., trend-based features, ratios)
4. Perform Initial Exploratory Data Analysis (EDA)
Once you shortlist variables, explore their relationship with the outcome:
- Correlation analysis: Identify which variables have a strong relationship.
- Feature distributions: Ensure variables are well-distributed and informative.
- Missing values check: See if you have enough data for each feature.
5. Feature Selection & Iteration
At this stage, you might:
- Use domain intuition to remove unlikely variables.
- Use statistical tests (e.g., mutual information, chi-square tests) to select features.
- Run a basic model (e.g., decision tree, logistic regression) to get initial insights.
Key Takeaways
- Start with a business problem, not data.
- Brainstorm potential drivers using domain knowledge.
- Validate data availability.
- Perform EDA to find patterns.
- Test and iterate.
Would you like to discuss this in the context of your specific fintech business case? I can help refine the approach further!
