Causal AI in Wealth, Explained

There’s a chart your mutual fund app shows you: “Investors who stay invested for 10 years earn 14% annually on average.” That chart is real. The advice is correct. And almost nobody follows it. Not because people are stupid — because knowing something and doing it are completely different problems. The entire wealth management industry has spent fifty years solving the wrong one. They’ve been obsessed with what to invest in. Nobody has seriously asked why people stop investing. India’s SIP stoppage ratio hit 109% in January 2025 — more SIPs cancelled than opened, for the first time in history. Causal AI is starting to answer why.

The Ice Cream Problem

In 1970s New York, city officials noticed something alarming: on days when ice cream sales spiked, so did the murder rate. The correlation was strong, consistent, and reproducible across years of data.

The actual explanation is summer. Hot weather causes people to buy ice cream. Hot weather also drives people outside, increases social friction, and causes more violence. Ice cream and murders share a common cause. Neither caused the other.

This is so obvious in retrospect that it’s almost funny. But the same mistake — confusing correlation with causation — is embedded in virtually every financial model ever built, every robo-advisor ever launched, and every “AI-powered” investment app you’ve ever used.[1]

Marcos López de Prado at Cornell and ADIA Lab showed with 10,000 Monte Carlo simulations that correlation-based factor models frequently produce portfolios that buy what they should sell and sell what they should buy. He called the result a “factor mirage.”

When a wealth platform builds its recommendation engine, it looks at historical data and asks: what patterns predict good outcomes? “Users who rebalance quarterly outperform.” “Investors who pause SIPs during corrections underperform.” All of this is correlation. The data is real. But correlation is not the same as causation, and the difference matters enormously when you’re trying to do something about it.

Correlation ≠ Causation: The Ice Cream & Murder Puzzle

Click “Reveal the Confounder” to see why ice cream doesn’t cause murder. The hidden variable — temperature — drives both.

In 1970s New York, ice cream sales and murder rates showed a strong correlation (r = 0.87). A scatter plot of monthly data reveals both variables rise and fall together. But neither causes the other. The confounder is temperature: hot weather causes both ice cream purchases and increased violence. When data points are colored by temperature, the pattern becomes obvious — cold months cluster at the bottom-left (low ice cream, low murder), hot months at the top-right. The correlation is real but the causal inference is wrong. This same mistake pervades financial modeling.

Here’s a concrete example. Suppose you observe that investors who check their portfolio frequently have better long-term returns. Correlation is clear. So you design a feature to make people check more. Does this work? Probably not. People who are financially committed both check their portfolio more and have better returns. The checking didn’t cause the returns. Both were caused by the same upstream thing: genuine financial engagement.

II.

Confounders & Colliders

Before going further, two concepts that explain exactly how correlations mislead you.

A confounder is a hidden third variable that causes both things you’re observing. Summer causes both ice cream and murders. Financial commitment causes both app usage and SIP continuation. If you don’t measure the confounder, you’ll incorrectly conclude one thing caused the other.

A collider is trickier. It’s a variable that is caused by two things you’re measuring. And when you try to control for it, you create a fake relationship between those two things that wasn’t there before.[2]

The classic example: shoe size and reading ability are unrelated. But both are caused by age. If you filter to only look at children of the same age, shoe size and reading ability suddenly look correlated — because you’ve controlled for the collider (age) and manufactured a relationship that doesn’t exist.

In finance, this happens constantly. Adding more variables to your model to make it look sophisticated often makes it worse — not because the variables are irrelevant, but because some are colliders, and including them corrupts everything else.

López de Prado’s 10,000 simulations. With 10,000 Monte Carlo trials, he showed that correlation-based factor models — the kind used by every major asset manager — frequently produce portfolios that buy what they should sell and sell what they should buy. He called the result a “factor mirage” — a pattern that looks real in backtests but evaporates with real money.

Confounder vs Collider: Two Ways Correlations Lie

Toggle between the two structures. A confounder hides above, creating a spurious link below. A collider sits below, and controlling for it creates a fake link above.

Confounders and colliders are the two specific ways correlations mislead causal reasoning. A confounder is a hidden common cause: in a triangle diagram, Summer sits at the top causing both Ice Cream Sales and Murder Rate below. The spurious correlation between ice cream and murder (r = 0.87) disappears when you control for Summer. A collider is the opposite structure: Financial Commitment and SIP Retention both cause App Engagement. When you filter on high-engagement users (controlling for the collider), you create a fake negative correlation between commitment and retention that doesn’t actually exist. Understanding which variables are confounders and which are colliders is essential for correct causal inference.

III.

Pearl’s Ladder

Judea Pearl — who won the Turing Award in 2011 for his work on causal inference — describes three levels of causal reasoning. They map almost perfectly onto the gap between what current wealth apps do and what’s actually possible.

Level 1: Observation. “What happened?” Your portfolio dropped 8%. Your SIP continuation rate is 70%. This is what every platform tells you.

Level 2: Intervention. “What happens if I do X?” If we send a goal reminder, what’s the probability they keep their SIP? This requires a causal model — you can’t answer intervention questions with pure correlation.

Level 3: Counterfactual. “What would have happened if?” You stopped your SIP in February 2022. If you had stayed invested, your portfolio would be ₹2.3 lakhs larger. This level of reasoning is only possible with a structural causal model.[3]

Pearl’s “Ladder of Causation” is formalized in his 2009 book Causality and popularized in The Book of Why (2018). The three levels correspond to increasing levels of causal reasoning that no amount of data alone can bridge.

Every current platform operates primarily on Level 1. Some gesture toward Level 2. Nobody has seriously built Level 3. But Level 3 is the one that actually changes behavior — because people don’t respond to abstract advice. “Stay invested for the long term” bounces off. But “if you had stayed invested through the last correction, you’d have ₹2.3 lakhs more right now” — that gets through.

Pearl’s Ladder of Causation: Three Levels

Each level requires strictly more causal knowledge than the one below. No amount of Level 1 data can answer Level 2 questions.

Judea Pearl’s Ladder of Causation has three levels. Level 1 (Observation) answers “What happened?” — showing that a portfolio dropped 8% and SIP continuation rate is 70%. Level 2 (Intervention) answers “What happens if I do X?” — predicting that a goal reminder would increase SIP continuation to 85%. Level 3 (Counterfactual) answers “What would have happened?” — calculating that staying invested through a specific correction would have yielded ₹2.3 lakhs more. Each level requires strictly more causal knowledge. Current platforms operate at Level 1. Level 3 is what actually changes investor behavior, because concrete counterfactuals are more persuasive than abstract advice.

The counterfactual is a mirror. And mirrors change people in a way that charts don’t.

IV.

The Causal Graph

Causal AI starts from a different question. Instead of “what patterns exist in this data?”, it asks “what actually causes what?” The machinery involves a causal graph — a diagram that explicitly maps which variables cause which other variables. Not just which ones correlate.

Here’s a causal graph for why an investor stops their SIP. A market drop triggers loss aversion, which leads to “I’m losing money, this was a mistake,” which forms a cancellation intent, and then — crucially — there’s a 21-day window before the SIP is actually cancelled.

A correlation-based system sees: “market down → SIP cancellations increase.” A causal system discovers: there’s a 21-day intervention window. And the right intervention isn’t market reassurance — it’s goal re-anchoring. Reminding the investor of the specific thing they’re investing for. When the goal becomes vivid again, short-term noise becomes irrelevant.

SIP Cancellation Causal Graph

Day after market drop: Day 0

Drag the slider to trace the causal chain. Then toggle “Goal Re-anchoring” to see how the right intervention in the 21-day window changes the outcome.

A causal graph for SIP cancellation shows a five-stage chain: Market Drop → Loss Aversion Activates → “I’m losing money” → Cancellation Intent Forms → SIP Cancelled. The key insight is a 21-day window between intent formation and actual cancellation. During this window, a goal re-anchoring intervention — reminding the investor of their specific financial goal — can reduce cancellation probability from 68% to 23%. This is fundamentally different from the industry’s standard response of sending market reassurance notifications, which have near-zero causal effect on cancellation behavior.

The intervention type matters, not the intensity. The industry sends market reassurance: “Markets are volatile. Stay the course.” Research on loss aversion shows this has near-zero effect. People in the grip of loss aversion are moved by something more primal — the vivid reality of what they’re trying to build. Their children’s education. Their freedom from financial anxiety. Reconnecting someone to their goal changes behavior in a way that market reassurance doesn’t.

The Engagement Trap

Here’s the specific mistake that most consumer fintech companies make. Engagement — how often users open the app, check their portfolio, read notifications — correlates with retention. Users who engage more tend to stay invested longer. So product teams optimize for engagement. More notifications. Gamification. Streaks. Confetti animations.

This doesn’t work. Engagement and retention share a common cause: financial commitment. People who are genuinely committed both engage more and stay invested longer. The engagement didn’t cause the retention.

In causal terms: engagement is a collider. Both commitment and retention cause engagement. When you control for engagement — by trying to maximize it — you open a spurious relationship. Your model thinks the nudges are working, but you’re really just measuring the committed users and calling the difference “impact.”[4]

The correct intervention is upstream. Not “get the user to engage more” but “increase the user’s financial commitment” — make the goal more concrete, vivid, and emotionally real. Engagement follows naturally.

The Engagement Trap: Collider Bias in Practice

The naive analysis shows engagement predicts retention (r = 0.71). The causal analysis reveals commitment drives both — engagement has zero causal effect on retention.

The engagement trap is a common collider bias in fintech. Naive analysis shows that high-engagement users have 82% SIP retention while low-engagement users have only 31% retention, suggesting a strong correlation (r = 0.71). Product teams conclude: increase engagement to increase retention. But causal analysis reveals that financial commitment is the common cause of both engagement and retention. When commitment is measured directly, engagement has zero additional causal effect on retention. The users who need help most are the quiet ones who stop engaging before cancelling. Optimizing for engagement selects for already-committed users while deprioritizing those who would benefit from intervention.

The users who actually need help are the ones who stop engaging. The quiet ones. The ones who don’t open the app for three weeks, then cancel their SIP on a Tuesday morning. Those users never received an intervention, because the engagement-optimization system had already classified them as “low-engagement” and deprioritized them.

VI.

The Counterfactual Mirror

What if, instead of a portfolio dashboard, your investment app showed you this: “You’ve started 3 SIPs in the last 4 years. You’ve stopped all 3. Here’s why — not what you might think.”

Each cancellation has a different root cause. One was pure loss aversion during a market dip. One was a calendar mismatch between SIP debit and salary credit dates. One was goal ambiguity — “general savings” was too vague to survive a competing priority. The total cost across all three: ₹74,400 in compounding.

Nobody has ever shown a user this. Because nobody has ever built the causal model required to produce it. This isn’t a feature. It’s a different kind of relationship between a product and a user — one built on accurate diagnosis rather than generic advice.

The Counterfactual Mirror: What If You Had Stayed?

Three SIP cancellations, three different root causes, three different counterfactuals. The causal model diagnoses each one differently.

The Counterfactual Mirror shows three real SIP cancellation scenarios with diverging paths. SIP 1 (February 2022): stopped during a Nifty downturn of 9.2% despite stable income, caused by pure loss aversion. The market bottomed 6 weeks later. Counterfactual cost: ₹38,400 in lost compounding. SIP 2 (August 2023): stopped during flat markets because the SIP debited on the 3rd but salary arrived on the 7th, creating 4 months of low-balance anxiety. Root cause was a calendar mismatch, not investment performance. Cost: ₹21,200. SIP 3 (January 2024): stopped with no market trigger, but 30 days after mentioning a car purchase with a vague “general savings” goal. Root cause was goal ambiguity. Cost: ₹14,800. Total cost of stopping: ₹74,400.

Why this is hard. Causal models require intellectual humility (you must commit to testable assumptions), longitudinal data (the same users observed over years), domain knowledge (you can’t learn causal direction from data alone), and new tooling (DoWhy, EconML, tigramite). All barriers are real. They’re also exactly why the competitive moat is deep.

Causal AI is not magic. It doesn’t predict the market. It doesn’t guarantee returns. What it does is give you an honest accounting of what’s actually happening — in your portfolio, in your behavior, in the interventions that do and don’t work. That’s the difference between knowing what you should do and actually doing it. Between advice and intervention. Between a chart and a mirror.

Based on the work of Judea Pearl (Causality, 2009), Marcos López de Prado et al. (Causal Factor Analysis, ADIA Lab, 2025), Susan Athey & Stefan Wager (Causal Forests, 2018), and Victor Chernozhukov et al. (Double Machine Learning, 2018). Market data from AMFI monthly SIP statistics.