Swiggy has built an internal artificial intelligence model that predicts the long-term value of new customers before they place their first order, and the company achieved a 63% reduction in model size while improving accuracy by adding a second prediction task. The Indian food delivery and quick commerce firm detailed its predicted lifetime value (pLTV) system in a recent technical post, explaining how the model enables the company to optimize advertising bids based on projected customer worth rather than immediate conversion metrics. The system addresses a core challenge in digital marketing: identifying high-value users when almost no behavioral data exists yet.
The model draws on more than 350 features spanning acquisition channel, device and fraud signals, geographic behavior, complaint history, Instamart category preferences, order behavior, payment patterns, and socioeconomic indicators—all generated from information available before a customer's first transaction. Swiggy's architecture uses a multilayer perceptron with three shared hidden layers that learn representations across prediction tasks, followed by separate heads for Food and Instamart predictions. By introducing order count as an auxiliary prediction task alongside lifetime value, the resulting four-head model shrank from 363,000 parameters to 135,000 while delivering marginally better results. The system achieved a Spearman correlation above 0.75 for both Food and Instamart, with 70% to 80% diagonal coverage across value deciles—meaning the model correctly ranked customers by actual value seven to eight times out of ten.
The prediction challenge stems from sparse early signals and a highly skewed target distribution, according to Swiggy, with a large share of new users placing no orders during their first 30 days while a relatively small group generates substantially higher lifetime value. Soumyajyoti Banerjee, staff data scientist at Swiggy, noted the counterintuitive result: "adding a second prediction task shrank the model by 63%—and made it more accurate." Jayshmi A, an engineer at Swiggy, explained that conventional regression metrics such as mean absolute error and mean absolute percentage error weren't sufficient for evaluation because of the large number of zero-value users and the long tail of customer value, so the company focused instead on whether the model could correctly rank customers by worth.
The company's internal model outperformed a third-party pLTV platform in production A/B tests, delivering higher retention and gross order value per acquired user without additional platform costs, though the external platform acquired more users per dollar spent. Swiggy deploys the pLTV signal with Google's target return on ad spend bidding, ranking customers by predicted value so the advertising system can differentiate bids accordingly. The report states the next phase will shift from point estimates to probabilistic pLTV predictions, aiming to represent uncertainty in customer value and incorporate it into bidding decisions. That evolution reflects a broader shift in how platforms balance model complexity against business impact—getting more insight from less computation by structuring the problem differently rather than simply adding capacity. For marketers managing acquisition costs across thin-margin businesses, the ability to predict value before conversion happens could reshape how platforms allocate their most expensive resource: attention.