--- title: Swiggy Uses 350+ Features and Multi-Task MLP to Predict... description: Swiggy built an in-house predicted lifetime value (pLTV) model to estimate the long-term value of new customers across its food delivery and Instamart quick... canonical: https://daily.dev/posts/swiggy-uses-350-features-and-multi-task-mlp-to-predict-customer-lifetime-value-sr2j36epj twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Swiggy Uses 350+ Features and Multi-Task MLP to Predict Customer Lifetime Value | daily.dev og:description: Swiggy built an in-house predicted lifetime value (pLTV) model to estimate the long-term value of new customers across its food delivery and Instamart quick... og:url: https://daily.dev/posts/swiggy-uses-350-features-and-multi-task-mlp-to-predict-customer-lifetime-value-sr2j36epj og:image: https://api.daily.dev/og/posts/sR2j36EpJ.png og:image:alt: Swiggy Uses 350+ Features and Multi-Task MLP to Predict Customer Lifetime Value og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Swiggy Uses 350+ Features and Multi-Task MLP to Predict Customer Lifetime Value **[InfoQ](https://daily.dev/sources/infoq)** · 3 min read · 0 upvotes · 0 comments ## Summary Swiggy built an in-house predicted lifetime value (pLTV) model to estimate the long-term value of new customers across its food delivery and Instamart quick commerce businesses, generating signals before a customer's first order. The model uses over 350 pre-order features and a multi-task MLP with shared hidden layers and task-specific heads for Food and Instamart. Adding order count as an auxiliary task cut parameters from 363,000 to 135,000 (a 63% reduction) while improving accuracy. Evaluation relied on decile-based ranking rather than standard regression metrics due to sparse, skewed data, achieving Spearman correlation above 0.75 and 70-80% diagonal coverage. The signal feeds into Google's target ROAS bidding, and Swiggy's in-house model outperformed a third-party pLTV platform on retention and order value in A/B tests, though the external platform acquired users more cheaply. Next steps involve moving toward probabilistic pLTV predictions. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much did adding order count as an auxiliary task reduce Swiggy's pLTV model size? Adding order count as a second prediction task alongside lifetime value reduced Swiggy's multi-task MLP parameter count from 363,000 to 135,000, a 63% reduction, while also marginally improving predictive accuracy. The architecture uses three shared hidden layers followed by task-specific heads for Food and Instamart. _Anyone tuning multi-task models for efficiency gains can track more real-world architecture writeups like this on daily.dev._ ### How do you evaluate a customer lifetime value model when most users have zero orders? Standard regression metrics like mean absolute error and mean absolute percentage error fail when a large share of users have zero value and a long tail skews the distribution. Swiggy instead divided predicted and actual values into ten quantiles and measured how often users landed in matching predicted and actual deciles, achieving a Spearman correlation above 0.75 and 70-80% diagonal coverage. _Data scientists wrestling with skewed target distributions can follow more evaluation techniques like this on daily.dev._ ### How many features does Swiggy use in its predicted lifetime value (pLTV) model? Swiggy's pLTV model uses more than 350 features, grouped across acquisition channel, device and fraud signals, geographic behavior, complaint history, Instamart category affinity, order behavior, payment patterns, and socioeconomic indicators. All features are generated from information available before a customer's first order to address the cold-start problem in advertising systems. _Engineers designing cold-start prediction pipelines can keep up with feature engineering approaches like this on daily.dev._ ## Similar posts on daily.dev - [Swiggy Improves Search Autocomplete Using Real Time Machine Learning Ranking](https://daily.dev/posts/swiggy-improves-search-autocomplete-using-real-time-machine-learning-ranking-u1igryh41) · InfoQ · 1 upvotes · 0 comments - [From Rigid Rules to Empathetic Agents: Rebuilding Swiggy’s Order Cancellation Experience with LLMs](https://daily.dev/posts/from-rigid-rules-to-empathetic-agents-rebuilding-swiggy-s-order-cancellation-experience-with-llms-9q9lewby1) · Swiggy Bytes · 4 upvotes · 0 comments - [Ordering Biryani Without Missing a Six: The Swiggy × JioHotstar Story](https://daily.dev/posts/ordering-biryani-without-missing-a-six-the-swiggy-jiohotstar-story-p3qvhxjkh) · Swiggy Bytes · 29 upvotes · 1 comments - [Building On-Device Predictive Autocomplete in React Native](https://daily.dev/posts/building-on-device-predictive-autocomplete-in-react-native-x1aus3xuo) · Swiggy Bytes · 0 upvotes · 0 comments - [Instamart’s Generative AI Forecasting: Scaling from Flat Forecasts to Deep Hierarchies](https://daily.dev/posts/instamart-s-generative-ai-forecasting-scaling-from-flat-forecasts-to-deep-hierarchies-sm44pwri5) · Swiggy Bytes · 2 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#feature-engineering](https://daily.dev/tags/feature-engineering) [View this post on daily.dev](https://daily.dev/posts/swiggy-uses-350-features-and-multi-task-mlp-to-predict-customer-lifetime-value-sr2j36epj) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Swiggy Uses 350+ Features and Multi-Task MLP to Predict Customer Lifetime Value","url":"https://daily.dev/posts/swiggy-uses-350-features-and-multi-task-mlp-to-predict-customer-lifetime-value-sr2j36epj","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/swiggy-uses-350-features-and-multi-task-mlp-to-predict-customer-lifetime-value-sr2j36epj"},"datePublished":"2026-09-02T14:01:18.153Z","dateModified":"2026-09-02T14:12:05.432Z","description":"Swiggy built an in-house predicted lifetime value (pLTV) model to estimate the long-term value of new customers across its food delivery and Instamart quick...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/46398875eb9e25c1f2d975ae411a8210?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/46398875eb9e25c1f2d975ae411a8210?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"InfoQ","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"InfoQ","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/afc3bced3e1e4b188dd9127017a60e0c","url":"https://daily.dev/sources/infoq"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/swiggy-uses-350-features-and-multi-task-mlp-to-predict-customer-lifetime-value-sr2j36epj","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,feature-engineering","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"InfoQ","item":"https://daily.dev/sources/infoq"},{"@type":"ListItem","position":3,"name":"Swiggy Uses 350+ Features and Multi-Task MLP to Predict Customer Lifetime Value"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/swiggy-uses-350-features-and-multi-task-mlp-to-predict-customer-lifetime-value-sr2j36epj#faq","mainEntity":[{"@type":"Question","name":"How much did adding order count as an auxiliary task reduce Swiggy's pLTV model size?","acceptedAnswer":{"@type":"Answer","text":"Adding order count as a second prediction task alongside lifetime value reduced Swiggy's multi-task MLP parameter count from 363,000 to 135,000, a 63% reduction, while also marginally improving predictive accuracy. The architecture uses three shared hidden layers followed by task-specific heads for Food and Instamart. Anyone tuning multi-task models for efficiency gains can track more real-world architecture writeups like this on daily.dev."}},{"@type":"Question","name":"How do you evaluate a customer lifetime value model when most users have zero orders?","acceptedAnswer":{"@type":"Answer","text":"Standard regression metrics like mean absolute error and mean absolute percentage error fail when a large share of users have zero value and a long tail skews the distribution. Swiggy instead divided predicted and actual values into ten quantiles and measured how often users landed in matching predicted and actual deciles, achieving a Spearman correlation above 0.75 and 70-80% diagonal coverage. Data scientists wrestling with skewed target distributions can follow more evaluation techniques like this on daily.dev."}},{"@type":"Question","name":"How many features does Swiggy use in its predicted lifetime value (pLTV) model?","acceptedAnswer":{"@type":"Answer","text":"Swiggy's pLTV model uses more than 350 features, grouped across acquisition channel, device and fraud signals, geographic behavior, complaint history, Instamart category affinity, order behavior, payment patterns, and socioeconomic indicators. All features are generated from information available before a customer's first order to address the cold-start problem in advertising systems. Engineers designing cold-start prediction pipelines can keep up with feature engineering approaches like this on daily.dev."}}]} ```