Swiggy's 135K-Parameter MLP Beats Third-Party pLTV With 350+ Features

Swiggy's machine learning team has published the architecture behind its predicted lifetime value (pLTV) system, a model that scores new customers across its food delivery and Instamart quick-commerce verticals before a single order is placed. The system feeds directly into Google's target return on ad spend (tROAS) bidding, replacing short-term conversion metrics with a forward-looking value signal. The headline finding: introducing a second prediction task cut the model's parameter count by 63% while simultaneously improving accuracy — a result that reinforces the case for architectural specificity over brute-force scaling.

Cold-Start Feature Engineering

The core challenge is informational poverty: the model must produce a reliable score with no order history. Swiggy's solution is a feature set exceeding 350 inputs, partitioned across acquisition channel, device and fraud signals, geographic behavior, complaint history, Instamart category affinity, order behavior, payment patterns, and socioeconomic indicators. Every feature is derived from pre-order signals, making the vector available at registration. The target distribution compounds the difficulty: a large share of new users place zero orders in their first 30 days, while a small cohort generates disproportionately high lifetime value. The model must separate those cohorts before any behavioral signal differentiates them.

Four-Head MLP: Shared Trunk, Auxiliary Tasks

Rather than training separate models per domain or reaching for a transformer-class architecture, Swiggy built a multilayer perceptron with three shared hidden layers feeding into task-specific prediction heads. The initial design had two heads — one for Food lifetime value, one for Instamart lifetime value. The team then added order count as an auxiliary regression task alongside each domain's value prediction, producing a four-head model.

The auxiliary task functions as a regulariser. Because order count and lifetime value are correlated but not identical targets, the shared trunk learns a richer latent representation than a single-task objective demands. The measurable consequence:

The 63% parameter reduction — from 363,000 to 135,000 — comes from the auxiliary task guiding the shared representation toward generalisation, allowing the network to achieve the same or better discriminative power with shallower effective capacity. Staff Data Scientist Soumyajyoti Banerjee described it directly: adding the second prediction task shrank the model and made it more accurate simultaneously.

Evaluation: Ranking Over Regression

Mean absolute error and mean absolute percentage error collapse under a zero-inflated, long-tailed target distribution. Swiggy engineer Jayshmi A designed an evaluation framework around ordinal ranking fidelity instead. Predicted and actual customer values are each bucketed into ten quantiles; the metric is diagonal coverage — the fraction of users whose predicted decile matches their actual decile. A Spearman correlation above 0.75 across both Food and Instamart confirms the model's rank ordering tracks real economic outcomes, which is precisely what a bidding system requires. The evaluation sidesteps the distortion zero-value users impose on standard regression metrics and directly validates the business objective: separating high-value users from low-value users at acquisition time.

This reflects a broader principle visible across production ML deployments — that pipeline architecture and evaluation design often drive more practical improvement than the choice of model family.

Production Results and Next Steps

In live A/B experiments comparing the in-house pLTV system against a third-party pLTV platform, Swiggy's model delivered higher retention and higher gross order value per acquired user without additional platform licensing costs. The external platform acquired more users per dollar spent, but the in-house model won on downstream quality metrics. In production, the pLTV score routes into Google's tROAS bidding as a bid signal based on predicted customer rank.

The team's next step is moving from point estimates to probabilistic predictions, with the intent to model uncertainty in customer value and propagate it into bidding decisions. Whether the sparse early-signal problem that already complicates point estimation becomes intractable when variance must also be estimated pre-order is the open question. That a 135,000-parameter MLP is already competitive with commercial alternatives gives Swiggy a meaningful foundation to build on.