Original research

REAL-TIME BIDDING ANOMALY DETECTION: AN ADAPTIVE FRAMEWORK FOR IDENTIFYING UNUSUAL BID REQUESTS IN PROGRAMMATIC ADVERTISING

Pages 199-210

10.61552/JAI.2026.04.003

Sujan Kumar Reddy Challa

• View Article
• PDF 649 K

Abstract Real-time bidding (RTB) coordinates automated transactions among publishers, exchanges, and demand-side platforms within millisecond-scale latency windows. The volume and dynamic nature of bid request streams make it difficult to detect abnormal activity such as bot-generated requests, spoofed inventory, invalid traffic, and collusion-based fraud. Rule-based approaches struggle here because valid bidding behavior varies widely across publishers, devices, geographies, and campaigns, producing excessive false positives or negatives. This paper proposes a context-aware framework for detecting anomalous bidding behavior in RTB ecosystems, combining statistical profiling, machine-learning anomaly scoring, and contextual behavioral baselines within a three-layer detection engine that fuses these signals into an interpretable, auditable risk score supporting feature-level attribution. The framework is implemented and evaluated on a reproducible synthetic RTB bid-request stream (250,010 requests, of which 15,010 are injected anomalies spanning bot-generated bursts, spoofed inventory, invalid traffic, and collusion rings), against static rule-based, single-method statistical, and unsupervised machine-learning baselines. At its critical operating threshold the framework achieves precision 0.895, recall 0.782 and F1-score 0.835 at a false-positive rate of 0.0065, detecting 99.9% of bot bursts and 99.1% of invalid traffic. A lower monitoring (suspicious-tier) threshold, sized to a 5% false-positive review budget, raises overall recall to 0.928 and surfaces 42.2% of spoofed-inventory requests, a class every baseline almost entirely misses at its own operating point (best baseline recall 19.1%). Mean scoring latency is 16.595 ms per request (p99 46.962 ms), within the bid-time decision window. Because no labeled production RTB stream is publicly available, the evaluation uses a documented synthetic generator with fixed seeds; all generator parameters are reported so the study can be reproduced.

Keywords: Real-time bidding, programmatic advertising, anomaly detection, ad fraud, adaptive baseline, machine learning.

Received: 21.06.2026. Revised: 23.08.2026. Accepted: 10.10.2026.