Uplift Modelling

A brokerage wants to send a "you've left cash uninvested — here's an idea" email and needs to decide who to send it to. A standard model trained to predict "will this client make a trade this week" and sent to the highest-scoring clients will mostly select clients who were going to trade anyway, email or not. The email gets credit for trades it did not cause. Worse, some clients might be irritated by unsolicited advice and trade less after receiving it — a genuinely negative effect that a plain response-prediction model cannot even represent, because it never models what would have happened without the email at all.

Uplift modelling predicts the individual causal effect of the action itself: not $P(\text{trade})$, but the difference between $P(\text{trade} \mid \text{email})$ and $P(\text{trade} \mid \text{no email})$ for that specific client. It is the marketing- and targeting-flavoured application of the same idea as CATE estimation — same machinery, framed around "who should we act on" instead of "what is the effect."

The four quadrants

Every individual falls into one of four unobservable groups, defined by how they'd behave under both scenarios:

• Sure things — would trade whether or not they get the email. The email is wasted on them.
• Lost causes — would not trade either way. The email is wasted on them too.
• Persuadables — would not trade without the email, but would trade because of it. These are the only clients where sending the email creates value.
• Sleeping dogs — would trade without the email, but the email annoys them enough that they don't. Sending to this group actively destroys value.

A plain response model that ranks by $P(\text{trade} \mid \text{email})$ cannot distinguish sure things from persuadables — both score high. Only a model that estimates the difference between treated and untreated outcomes can single out persuadables, and only that model can spot sleeping dogs at all, since a response model never even computes the untreated baseline for a client it always shows the treatment condition to.

Worked example

Run a randomised test: half of 10,000 dormant clients get the email, half don't, and track who trades within a week. Suppose 30% of the emailed group trades versus 22% of the control group — an overall uplift of 8 points, which looks worth scaling to the full client base. But an uplift model conditioning on account features finds this 8-point average splits sharply: clients who logged in within the past month show +22 points uplift (strong persuadables), while clients dormant for over a year show −3 points (a small sleeping-dogs effect — the email seems to remind them the account exists and they let it lapse further, or unsubscribe and disengage entirely). Targeting only the recently-active segment captures nearly all the value the campaign can create while avoiding a segment where the campaign is actively counterproductive — a decision the aggregate 8-point number could never have revealed.

Where quants and quant-adjacent teams meet this

Beyond marketing emails, the same structure applies to deciding which accounts get a fee waiver to prevent churn, which clients get proactive risk-limit calls, or which counterparties get relationship-manager outreach before a renewal — any setting where an action is costly, has a plausible negative side, and a randomised or quasi-randomised holdout exists to measure it against.

Evaluating an uplift model without ground truth

The single hardest practical problem in uplift modelling is that you can never observe the ground truth for any individual — nobody is simultaneously emailed and not emailed, so there is no row anywhere with a directly observed uplift value to check a prediction against. The standard workaround is the uplift curve (sometimes called a Qini curve): rank all individuals in a held-out randomised test by predicted uplift, then walk down the ranking and compare, at each point, the actual observed difference in outcome rate between the treated and control individuals seen so far. A model that ranks persuadables correctly produces a curve that rises steeply at the top of the ranking (the highest-predicted-uplift individuals show the largest real treated-vs-control gap) and flattens or falls toward the bottom — the model-free equivalent of a lift curve, adapted to work with a causal rather than a predictive target.

Turning ranked uplift into a targeting decision

A ranked list of predicted persuadables is not yet a targeting rule — that requires weighing the cost of the action against its expected benefit. If sending the email costs very little, targeting everyone above a low uplift threshold, or even everyone except clearly-predicted sleeping dogs, can be worthwhile even with a noisy model. If the action is expensive (an outbound phone call, a fee waiver with real revenue cost), only the individuals with the highest, most confidently estimated uplift clear the bar. This mirrors the cost-sensitive framing used for ordinary classification decisions in Decision Thresholds and Cost-Sensitive Learning — the model produces a ranked estimate, but where to draw the line depends on numbers the model itself does not know.