Advertising Inventory Forecasting Using an LSTM-Based Deep Learning Model

The iQIYI advertising algorithm team proposes a deep‑learning algorithm based on LSTM (Long Short‑Term Memory) to forecast advertising inventory. The method aims to reduce training resources, improve model generalization, and incorporate fine‑grained temporal labels (workdays, holidays, etc.) to capture holiday‑induced inventory fluctuations.

Background : Advertising inventory forecasting is a time‑series prediction task crucial for ad strategy, requiring early estimation of user traffic. It faces three main challenges: (1) limited daily samples leading to potential over‑fitting, (2) strong holiday effects that cause large inventory spikes, and (3) multi‑dimensional targeting (site, platform, city, channel) that increases data sparsity and computational cost.

Traditional approaches such as moving average, exponential smoothing, and ARMA/ARIMA models struggle with non‑stationary data and cannot easily incorporate external features like holidays. The team previously used an Adaptive‑ARIMA (A‑ARIMA) model, which adds missing/abnormal value handling and holiday bias correction, achieving a 1.16 percentage‑point reduction in bias compared with standard ARIMA.

Proposed LSTM solution : The LSTM network uses a sliding window of the past N days as training samples, randomly sampled to avoid sequential bias. Key optimizations include:

Data normalization: maximum‑value scaling is chosen over min‑max to preserve the correlation of time‑series data across dimensions with vastly different scales.

Dimension clustering: Integrated Absolute Error (IAE) is computed between inventory curves, and hierarchical clustering groups dimensions with similar trends. Each cluster shares the same LSTM parameters, balancing generalization and resource usage.

Holiday features: a fully‑connected layer concatenates LSTM outputs with encoded holiday categories (regular workday, pre‑holiday, holiday, weekend), allowing the model to learn holiday‑driven patterns.

Experimental results : The LSTM model and A‑ARIMA were evaluated on 49 days of data from June to September 2019, covering both holiday periods and regular weeks. Using weighted bias rate as the metric, LSTM reduced the dimension‑weighted bias by 0.72 percentage points and the overall inventory bias by 1.48 percentage points per day, demonstrating superior forecasting accuracy.

Conclusion and outlook : The LSTM‑based inventory forecasting algorithm outperforms the A‑ARIMA baseline, thanks to effective normalization, dimension clustering, and holiday feature integration. Future work will explore richer holiday representations and further refinements to enhance prediction precision.