When Every Minute Is a Brand Promise: Swiggy's Delivery Time Prediction Algorithm

Industry & Competitive Context

India's online food delivery sector is among the most concentrated platform markets in Asia. As documented in Swiggy's DRHP filed with the Securities and Exchange Board of India (SEBI) in September 2024, the Indian food delivery market is effectively a duopoly, with Swiggy and Zomato together commanding over 90% of market share. Industry experts quoted in the same filing project the market's valuation to reach INR 2 lakh crore by 2030.

Against this backdrop, platform differentiation on traditional levers — restaurant variety, pricing, and discounts — has reached structural limits. Both platforms serve similar restaurant inventories, operate comparable loyalty programmes, and compete for the same pool of delivery executives (DEs). What remains meaningfully differentiable, and what customers experience on every single order, is the delivery time promise: the estimated time of arrival (ETA) displayed before checkout and updated live on the tracking screen. In a market where consumers routinely evaluate the ETA figure before confirming an order — and where even a few minutes' difference can shift a decision — the accuracy, credibility, and behavioural smoothness of the ETA number became a genuine marketing asset. Swiggy recognised this dynamic early and systematically invested in machine learning to convert ETA from a rough approximation into a precision instrument.

Brand Situation Prior to the Algorithm's Development

Swiggy was founded in 2014 in Bengaluru by Sriharsha Majety, Nandan Reddy, and Rahul Jaimini. By the time it prepared for its IPO — which saw the company file its DRHP with SEBI in September 2024 — Swiggy had scaled to over 150,000 restaurant partners and 112.7 million transacting users (as of June 2024). Its consolidated operating revenue for FY2024 stood at INR 11,247.4 crore, representing a 36% year-on-year increase. Net losses narrowed significantly, declining by 43% to INR 2,350.24 crore from INR 4,179.30 crore in FY2023. However, this scale created an operational paradox. The very growth of the platform — more orders, more restaurants, more cities, more delivery executives — made accurate time estimation increasingly difficult. As documented in Swiggy's official engineering blog ("The Swiggy Delivery Challenge, Part One," bytes.swiggy.com, 2019), even early-stage operations revealed that naively averaging historical delivery data produced systematically unreliable estimates. An average-based system, by mathematical definition, disappoints the lower half of all customers with optimistic estimates while discouraging potential orderers with conservative ones. Swiggy's technical team also identified a critical infrastructure gap: the default industry solution — routing APIs from commercial mapping providers — was inadequate for their operational reality. As stated in the same engineering blog post, commercial mapping APIs were too slow for the number of route combinations that needed to be calculated at scale, prohibitively expensive at Swiggy's order volumes, and structurally unable to model two-wheeler travel dynamics. Critically, such APIs could not account for what Swiggy's engineers called the "last last mile" — the time a delivery executive spends navigating from a society gate to the customer's door — a segment that can add five to ten minutes and that disproportionately affects food quality upon arrival. The brand implication was clear: inaccurate ETAs eroded customer trust, inflated customer-care contact rates, and contributed to cancellations. Building a proprietary, precision prediction system was therefore simultaneously an engineering priority and a brand reliability investment.

Strategic Objective

Swiggy's stated objective, as articulated across its engineering blog series, was to calculate what its engineers described as the "Goldilocks value": an ETA high enough to avoid disappointing customers with over-optimistic promises, but low enough not to discourage order placement. This is a fundamentally marketing-aligned objective — it seeks to manage consumer expectations with enough precision to maximise both trust and conversion. The system needed to serve two distinct consumer touchpoints simultaneously. First, the pre-order (cart) stage, where customers see an SLA (Service Level Agreement) estimate before placing an order, and which directly influences their decision to proceed. Second, the post-order tracking stage, where customers monitor live ETA updates and where unexpected "bumps" — sudden upward jumps in estimated delivery time — create anxiety and erode confidence in the platform. Designing a system that performed well at both stages, with different information available at each, required a fundamentally different architecture from a single prediction model.

Algorithm Architecture & Technical Execution

Swiggy's engineering team formalised the delivery time problem with a mathematical decomposition. As published in "The Swiggy Delivery Challenge, Part One" (bytes.swiggy.com, 2019), the delivery timeline follows a structural equation: Delivery Time = Max (Assignment Delay + First Mile Time, Prep Time) + Last Mile Time The Max() function reflects an important operational reality: food preparation at the restaurant and the dispatch of a delivery executive happen in parallel. The binding constraint is whichever takes longer, not the sum of both. This seemingly simple formula encodes a non-trivial insight — that assignment efficiency and kitchen throughput interact dynamically, and both must be modelled together. The production system breaks delivery time prediction into five distinct target variables, as documented in "Predicting Food Delivery Time at Cart" (bytes.swiggy.com, August 2023, authored by Shubham Grover, Soumyajyoti Bannerjee, Vaibhav Agarwal, Sunil Rathee, and Akshita Sood):

The MIMO Model Architecture

To predict all five variables simultaneously and exploit the mathematical relationships between them (O2R = O2A + FM + WT + LM), Swiggy's team deployed a Multi-Input Multi-Output (MIMO) deep learning model supplemented with entity embeddings, as detailed in the August 2023 blog post. The strategic logic of MIMO is that the five outputs share substantial overlapping feature inputs; training them jointly in a single network allows the model to learn cross-leg dependencies that separate models cannot capture. For instance, O2A and FM times are intricately linked because the assignment algorithm attempts to dispatch a DE so that they arrive at the restaurant precisely when food is ready — making the two variables co-determined rather than independent. Entity embeddings were used to handle high-cardinality categorical variables — specifically restaurant identifiers, customer geo-hashes, and city identifiers. Rather than one-hot encoding these into sparse, high-dimensional vectors, embeddings compress them into dense, semantically meaningful representations. This design choice allows the model to learn, for example, that two restaurants with similar preparation-time behaviour are proximate in embedding space, enabling generalisation to unseen order combinations.

Key Input Features

The documented inputs to the cart-stage prediction model include: the type of restaurant (cloud kitchens are noted to have faster preparation times than traditional dine-in establishments); the number and variety of items in the order (more items, or items that are inherently slower to prepare, extend prep time); current restaurant "stress" as derived from the ratio of orders placed to orders prepared in recent periods (a signal of kitchen congestion); availability of delivery executives in the hyperlocal zone; historical first-mile and last-mile speed patterns around the restaurant and customer locations; and near-real-time speed patterns derived from DE GPS pings.

Post-Order Tracking: The Four-Stage ETA System

The cart-stage model generates the initial SLA promise. Once an order is placed and progresses through fulfilment, a separate but complementary system takes over. As documented in "Where is my order? — Part I" and "How ML Powers — When is my order coming? — Part II" (bytes.swiggy.com, May 2023, authored by Vaibhav Agarwal), the post-order tracking ETA is managed by a four-stage architecture corresponding to the four order legs: O2A, FM, WT, and LM. Each stage uses different inputs that are only available once the order has progressed to that stage. For example, the FM stage can incorporate information about the specific delivery executive assigned — their familiarity with the restaurant location, their mode of transport (motorcycle, bicycle, or other), and their historical delivery patterns — inputs that are unavailable at cart-stage. At the LM stage, real-time DE GPS pings are used to compute derived features including haversine distance to the customer, cumulative distance travelled (which more accurately captures actual road distance than a straight-line calculation), and global DE speed — all of which serve as real-time traffic proxies. A secondary metric introduced alongside the standard MAE (Mean Absolute Error) was the "Percentage of Orders with Inaccurate Bumps" — the proportion of orders where the estimated ETA exhibited sudden, misleading upward or downward jumps. This metric directly captures the consumer-experience dimension of ETA quality; smooth, gradually improving estimates are less anxiety-inducing than accurate but volatile ones.

The Evolution from Gradient Boosting to Neural Networks

Swiggy's ETA system underwent a documented architectural evolution. The engineering team initially used Gradient Boosting Trees (GBT) for the last-mile estimation, but noted — as recorded in the May 2023 Part II blog — that GBT models are static once trained and do not adapt to dynamic, real-time data changes. The team transitioned to a neural network architecture: four hidden layers per model, each using Leaky Re LU activation functions, trained for 50 epochs using the ADAM optimiser. This shift was motivated by the need for greater adaptability to real-time signals — particularly live GPS pings and current traffic states — which change continuously during an order's journey.

Positioning & Consumer Insight

The central consumer insight animating Swiggy's ETA investment was explicitly stated in its engineering blog: "Based on the performance of our estimates, customers tend to form a perception of the accuracy of these promises. So accurate delivery time predictions are also important for inspiring customer trust in Swiggy." (Swiggy Bytes, August 2023.) This is a textbook positioning insight: the ETA number is not merely a logistics output — it is a trust signal. Each time an order arrives within the promised window, the brand deposits into its credibility account. Each time it does not, a withdrawal is made. The insight also carries a conversion dimension. As articulated in "The Swiggy Delivery Challenge, Part One" (2019): "If we show too low a delivery time and we don't deliver within it, it leads to bad customer experience. [But] every increased minute results in fewer people ordering." This is an explicit acknowledgment that the ETA estimate functions as a demand lever — it directly affects top-of-funnel order intent, not just post-order satisfaction. Accordingly, the engineering problem and the marketing problem are structurally identical: find the precise, credible estimate that maximises both trust and conversion simultaneously.

Platform & Channel Strategy

The ETA prediction system operates as an embedded capability within Swiggy's consumer app (iOS and Android) and web platform — the primary channels through which customers place orders. The estimate appears at two strategically important friction points: the restaurant listing page (where a customer decides whether to order from a given restaurant), and the checkout/cart page (where the final order confirmation occurs). Post-order, the live tracking screen continuously refreshes the ETA, functioning as a real-time brand promise update. No verified public information is available on specific A/B testing protocols, ETA display design decisions (e.g., whether ranges versus point estimates were tested), or channel-specific marketing communications built around ETA accuracy claims. The DRHP filed in September 2024 identifies technology and cloud infrastructure investment as an intended use of IPO proceeds, confirming that the platform capability continues to be treated as a strategic investment area, but does not detail specific channel deployment decisions.

Business & Brand Outcomes

Swiggy's engineering team published the following verified performance improvements, drawn directly from the official Swiggy Bytes blog posts: Beyond these engineering metrics, the Swiggy Bytes Part II blog (May 2023) documented a directional qualitative outcome: the revamped ETA model architecture produced "a notable decline in customer-care agent interactions and order cancellations." No specific numerical values for this decline were published in the blog or in any other verified public source.

Strategic Implications

Proprietary Data as Competitive Moat. Swiggy's decision to build a bespoke ETA system rather than rely on commercial mapping APIs reflects a core strategic principle: at platform scale, first-party operational data becomes a proprietary asset that off-the-shelf tools cannot replicate. The ability to model two-wheeler dynamics, apartment-complex last-mile behaviour, and restaurant kitchen stress in real time is only achievable with years of proprietary delivery data. This creates a compounding moat — as the platform grows, the model improves, which improves the product, which attracts more users, which generates more training data.

Decomposition as an Engineering and Strategic Discipline. Swiggy's explicit decomposition of a single "delivery time" number into five granular sub-predictions (O2A, FM, WT, LM, O2R) reflects a strategic discipline that extends beyond engineering. By isolating each leg of the journey, the company gains diagnostic visibility into precisely where delays occur — enabling targeted operational intervention. From a management perspective, this decomposition converts a consumer experience problem (the ETA is wrong) into five specific operational problems, each with an accountable owner.

The MIMO Architecture as an Analogy for Integrated Strategy. The choice of MIMO over separate, siloed models encodes a management insight: outputs that are mathematically interdependent should be optimised jointly, not independently. This principle maps directly onto marketing strategy, where brand promise, operational execution, and customer communication are interdependent variables that degrade when managed in silos.

ETA as a Brand Contract, Not a Forecast. Perhaps the deepest strategic implication is definitional. Swiggy's engineering blog frames the ETA not merely as a prediction but as the mechanism through which "customers form a perception of the accuracy of these promises." This reframing — from output to contract — has significant implications for how product, technology, and marketing teams should be organised and incentivised. If ETA accuracy is a brand metric, it belongs on the same dashboard as NPS and satisfaction scores, not just in the engineering team's MAE reports.

Technology Disclosure as Brand Signal. Swiggy's decision to publish detailed technical explanations of its ETA system on its official engineering blog (bytes.swiggy.com) is itself a strategic communications choice. Publishing the methodology — including the shift from GBT to neural networks, the introduction of entity embeddings, and the specific performance gains achieved — signals sophistication to potential talent, to restaurant partners evaluating platform reliability, and to investors assessing technology depth ahead of the IPO. Technical transparency, selectively deployed, functions as a form of competitive credibility marketing.

Discussion Questions

1. Swiggy's engineering documentation frames ETA accuracy as a driver of customer trust and a conversion lever. Using the concept of the "service promise" from service marketing theory, how would you position ETA accuracy within Swiggy's overall brand equity framework? What is the risk of optimising ETA accuracy in isolation from actual delivery performance?

   
   

2. The MIMO architecture was chosen over separate models because the five delivery-time components are mathematically interdependent. What organisational analogue does this architectural choice present for marketing strategy? In what strategic contexts does "joint optimisation" of interdependent outcomes outperform the management of each outcome independently?

   
   

3. Swiggy explicitly chose not to use commercial mapping APIs, citing cost, speed, and two-wheeler modelling limitations. Evaluate this build-versus-buy decision using a resource-based view of strategy. Under what conditions should a platform company treat a logistics intelligence capability as a core competency versus an outsourced function?

   
   

4. The metric "Percentage of Orders with Inaccurate Bumps" was introduced to capture the consumer anxiety created by sudden ETA revisions — even when the final delivery time might still be accurate. What does this metric reveal about the relationship between statistical accuracy and perceived service quality? How should companies design measurement frameworks to capture both dimensions?

   
   

5. Swiggy's public disclosure of detailed ETA modelling methodology on its engineering blog is a form of strategic communication. Analyse this disclosure decision through the lens of competitive strategy: what are the benefits and risks of publishing technical differentiation, and under what conditions is such transparency strategically rational versus self-defeating?