Applying PID Control Algorithm for Advertising Cost Control in Real-Time Bidding Systems

In a highly competitive market where advertising is a crucial channel for brand communication, internet ad platforms face the challenge of balancing cost and return, especially when multiple advertisers bid for the same ad slot.

To address cost‑control issues, two main approaches are used: manual real‑time monitoring and algorithmic automation. After evaluating various methods, the PID (Proportional‑Integral‑Derivative) control algorithm was selected for its simplicity, broad applicability, and ease of deployment.

PID Control Algorithm Overview

The PID controller, widely used in industrial control, consists of three terms: proportional (P), integral (I), and derivative (D). It computes an output based on the error between the measured value and the target, using the formula shown in the image below.

The three terms have distinct effects:

Proportional term : reacts instantly to error, reducing deviation quickly but cannot eliminate steady‑state error; excessive gain leads to overshoot and oscillation (see Fig. 2‑2).

Integral term : accumulates error over time to eliminate steady‑state error; too large a gain can cause large overshoot and instability (see Fig. 2‑3).

Derivative term : predicts error trend, providing early correction to improve response speed and reduce overshoot; excessive derivative gain may destabilize the system (see Fig. 2‑4).

In practice, a PI controller is often used first to simplify tuning, then the D term is added if needed.

Practical Case: Information‑Flow Advertising

Advertisers purchase media slots and aim to keep the transaction price close to a pre‑planned budget. Because bidding models and second‑price mechanisms introduce uncertainty, direct price control is impossible; instead, the bid is adjusted dynamically.

The PID feedback loop for bid adjustment is illustrated below:

Since the initial bid cannot be zero, the raw PID output is treated as a delta and added to the previous bid, preventing large early fluctuations.

Historical bid‑price data were analyzed, revealing a positive correlation with random noise. A hybrid model combining a linear regression (minimum‑MSE fit) and a stochastic component based on the empirical distribution was built to simulate realistic transaction prices.

Parameter tuning experiments showed that:

Large proportional gain caused significant oscillations.

Large integral gain delayed convergence to steady state.

Large derivative gain produced persistent oscillations.

Figures 3‑9 to 3‑12 compare fixed‑bid strategies with the PID‑adjusted strategy under two target‑cost scenarios (A and B). The PID approach keeps the average transaction price close to the target, while the fixed‑bid method often deviates, especially over short campaign periods.

Finally, the PID controller was refined to respect bid limits and timing constraints, and combined with other ad‑system components (CTR/CVR prediction, audience segmentation, budget allocation) to achieve overall optimal ad performance.

Conclusion

PID control, originally designed for industrial processes, can be successfully adapted to online advertising cost control. After modest algorithmic adjustments and offline parameter tuning using simulated price data, the PID‑based system demonstrated stable budget adherence and improved cost efficiency compared with static bidding strategies.