Api Response Time Prediction Using a Deep Learning Model on Web Application Load Testing Simulation Data

Abstract

Web application performance, particularly application programming interface (API) response time, is one of the primary determinants of user experience and system reliability, especially under high-load conditions. Conventional load testing is reactive, as it can only measure actual response time after a given load has been applied to the system, creating a need for a predictive approach capable of estimating API response time from load and server-resource conditions. This study applies a deep learning model in the form of a Multi-Layer Perceptron (MLP) with three hidden layers to predict API response time, and compares its performance against two baseline models, namely Linear Regression and Random Forest Regressor. Due to limited access to sensitive real-world production data, this study uses a simulated load-testing dataset of 500 samples generated programmatically, comprising ten load and server-resource features such as concurrent users, requests per second, payload size, database query count, cache hit ratio, and CPU and memory utilization, with non-linear patterns representing resource-contention effects under high load. The data was split into 70% training and 30% testing, then evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²). Experimental results show that the Deep Learning (MLP) model achieved the best predictive performance with an RMSE of 24.86 ms, MAPE of 19.33%, and R² of 0.9009, outperforming Linear Regression (RMSE 26.25 ms, R² 0.8895) and Random Forest (RMSE 27.90 ms, R² 0.8751), although Linear Regression's MAE was slightly lower (19.34 ms versus 19.53 ms). Feature-importance analysis indicates that the number of concurrent users, database query count, and CPU utilization are the most dominant factors influencing API response time. To ground the simulated scenario in a concrete example, a small real demonstration web application was also built and load-tested; its qualitative behavior (response time increasing with concurrency) is consistent with the assumptions underlying the simulated dataset. These findings indicate that the deep learning model is better able to capture non-linear patterns caused by resource contention compared to a linear model or a tree-based ensemble, suggesting potential use as a predictive component in auto-scaling strategies or web-application capacity planning, with the caveat that these results remain a preliminary study requiring further validation using real production data.