AI-Driven Sales Forecasting in the Gaming Industry: Machine Learning-Based Advertising Market Trend Analysis and Key Feature Mining

1. Introduction

The game industry is undergoing a data-driven marketing transformation, where advertising plays a vital role in user acquisition and revenue growth. Existing studies have explored time series forecasting based on sales records, click-through rate (CTR) modeling, and conversion behavior prediction. For example, some works apply ARIMA and LSTM to forecast sales, while others use gradient boosting to model ad performance metrics. However, few approaches effectively combine multi-source advertising features with trend dynamics in a unified prediction framework. Furthermore, current methods often overlook the impact of time-lagged variables and do not fully explain the relative importance of features. In response to these gaps, this paper proposes a hybrid prediction model that integrates XGBoost and LSTM to capture both static behaviors and temporal patterns. By incorporating trend features, SHAP-based interpretability, and a dynamic feature window mechanism, the model enhances robustness and interpretability. The core innovation lies in aligning causal modeling with advertisement-driven sales changes, offering a comprehensive, interpretable, and scalable solution for sales forecasting in the game industry.

2. Data Source and Feature Construction

2.1. Data Collection and Preprocessing Process

The experimental data is provided by a mobile game company, including advertising logs, user conversion data and sales records, covering a period of 180 days. The advertisement log covers basic fields such as placement time, platform channel, advertisement position type, creative material logo, exposure volume, click volume, and cost expenditure. Sales data are daily downloads and activations, which are used as target variables for model training.

In the data preprocessing stage, a uniform time granularity (in days) is constructed according to the field types, missing items and outliers are handled, and the de-emphasis of ad logs and standardized coding of materials are carried out. At the same time, the metrics of different platforms are standardized to a uniform scale to ensure comparability of features. The time field is uniformly converted to a relative timestamp to provide a basis for subsequent lag modeling and sliding window feature construction.

2.2. Advertising Indicator System and Trend Variable Extraction

The advertising feature system is constructed based on domain knowledge and statistical screening, covering placement behavior, user response, and cost-effectiveness. Feature selection combines correlation analysis and domain relevance: Pearson coefficients and lag response studies were used to quantify relationships with sales, while practical campaign insights guided the inclusion of variables such as MaterialAge and budget ratio. All categorical features are encoded numerically, and continuous variables are transformed into daily time series aligned with the sales timeline. Trend variables—such as moving averages and volatility rates—are derived from base metrics to enhance temporal sensitivity in forecasting models.

Table 1. Structure of advertising indicator system design.

Dimension	Indicator Name	Type	Meaning
Placement Behavior	Budget	Continuous variable	Daily budget (yuan)
	Impressions	Continuous variable	Daily Exposure (times)
	Platform	Categorical Variable	Platform Code
	AdSlot	Categorical Variable	AdSlot Type
User Response Class	Clicks	Continuous Variables	Clicks on Ads
	CTR	Continuous Variable	Click Through Rate (Clicks/Exposure)
	CVR	Continuous Variable	Conversion Rate (Activation/Click)
Cost Effectiveness Category	CPM	Continuous Variable	Cost Per Thousand Displays ($)
	CPC	Continuous variable	Cost per click ($)
	CPI	Continuous Variable	Cost per installation (yuan)

Table 1. Structure of advertising indicator system design.

Dimension	Indicator Name	Type	Meaning
Placement Behavior	Budget	Continuous variable	Daily budget (yuan)
	Impressions	Continuous variable	Daily Exposure (times)
	Platform	Categorical Variable	Platform Code
	AdSlot	Categorical Variable	AdSlot Type
User Response Class	Clicks	Continuous Variables	Clicks on Ads
	CTR	Continuous Variable	Click Through Rate (Clicks/Exposure)
	CVR	Continuous Variable	Conversion Rate (Activation/Click)
Cost Effectiveness Category	CPM	Continuous Variable	Cost Per Thousand Displays ($)
	CPC	Continuous variable	Cost per click ($)
	CPI	Continuous Variable	Cost per installation (yuan)

2.3. Engineering Game Sales Related Features

Sales prediction not only relies on advertising indicators, but also needs to construct upstream and downstream features related to sales response. User feedback type variables are constructed, including the average number of activations in the last 7 days (Activation-MA7), the estimated value of retention rate (RetRate-Est), and Material-Age variables are introduced to portray the timeliness of advertisement creativity. In order to enhance the feature expression capability, cross-feature combinations are introduced, such as "ad platform × ad space type" to capture platform differentiation. In addition, indicators such as budget share and material repetition are normalized and embedded as features in the model to expand the interpretable dimensions of the variables. The feature set is encoded uniformly by One-Hot, Z-score, Log-Transform and other processes for input modeling.

2.4. Feature Time-Series Relationship Analysis

There is a significant time-series dependence between advertising data and sales variables, and it is necessary to analyze the time response pattern of various types of features to sales at the pre-modeling stage, in order to construct appropriate lag windows and trend variables. The lagged versions of the original advertising indicators are constructed by means of sliding windows, including the feature versions under the granularity of Lag-1, Lag-3, Lag-7, etc., which are used to simulate the potential delayed effects of different variables on sales. On this basis, sliding average (Moving Average), weighted decreasing average (EWMA) and other methods are introduced to construct the stability trend features, which are used to reflect the smoothness and direction of change of variables in the time dimension.

3. Prediction Model Design and Training Methods

3.1. Overall Architecture and Technical Route

Based on the AI-driven sales volume prediction task, the model architecture takes advertising market features as the core input, integrates behavioral variables and trend variables, and constructs a multi-source fusion prediction system. The overall structure includes an input layer, a feature fusion layer, a core modeling layer and an output prediction layer, which comprehensively responds to the modeling needs of "advertising market trend analysis and key feature mining". As the core input source, advertising data is mapped into a high-dimensional input tensor through unified time-series coding and trend window conversion. Combined with the sales target variables, a regression prediction model is constructed, and the objective function is the weighted mean square error:

$$L={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{w}_i\cdot {(y_i-{\widehat{y}}_i)}^2}$$

(1)

where $L$ is the overall loss value,$n$ is the total number of samples,$w_i$ is the weight of the$i$ th sample,$y_i$ is the real sales, and${\widehat{y}}_i$ is the forecast sales. The weight$w_i$ is used to strengthen the influence of high-budget material samples in learning, reflecting the weight orientation of advertising resources. The overall modeling process is shown in Figure 1, where all the input variables are designed around the changes in the state of the advertising market, and mapped to the sales output through the AI model, realizing the non-linear causal modeling path from "advertising behavior" to "sales result".

3.2. Basic Model Selection

Considering that the advertising behavior variables are highly sparse, strongly time-dependent, and have drastic trend dynamics, the model adopts the artificial intelligence modeling strategy that synergizes the static structure with the temporal structure. The baseline adopts XGBoost to process advertising behavior and static features, with excellent adaptability to high-dimensional discrete features, and its update process is as follows:

$${\widehat{y}}_i^{(t)}={\widehat{y}}_i^{(t-1)}+\eta \cdot f_t(x_i)$$

(2)

where$f_t$ is the$t$ th regression tree, capturing the marginal impact of advertising metrics on sales. A large number of trend variables in advertising market features are strongly time-series, modeled by LSTM structure, and its hidden state is updated as:

$$h_t=\sigma (W_h{x}_t+U_h{h}_{t-1}+b_h)$$

(3)

where $h_t$ is the hidden state at the current moment,$x_t$ is the input vector,$h_{t-1}$ is the hidden state at the previous moment,$W_h、U_h$ is the weight matrix,$b_h$ is the bias term, and$\sigma$ is the activation function. LSTM is specially designed to learn the nonlinear process of the response of advertisement variables (e.g., CTR, CVR, etc.) to the sales volume over time. In order to improve the semantic coverage of the model, a hybrid LSTM-MLP structure is constructed to realize the joint expression of behavioral variables and trend variables, and complete the deep feature extraction of "key influencing factors" in the advertising data. Table 2 summarizes the input type, main structure and fitting ability of the three types of models.

3.3. Multi-Source Feature Fusion Modeling Approach

In order to realize the goal of "key feature mining" in the question, a multi-source data fusion strategy is used to integrate the advertising data. The advertisement behavior feature matrix is denoted as$X_c\in R^{n\times d_c}$ , and the embedded classification variable is$X_d\in R^{n\times d_d}$ . The static features and embedded features are jointly expressed as:

$$X_{input}=\left[X_c‖\mathrm{Embed}(X_d)\right]$$

(4)

where $X_{input}$ is the final model input features,$X_c$ is the continuous feature matrix,$X_d$ is the discrete feature set, and$\mathrm{Emded}(\cdot )$ denotes the embedding operation.

The advertising market trend features are generated by sliding time window to generate a 3D tensor$T\in R^{n\times w\times k}$ , which is inputted into LSTM to get the time-series representation$h_T$ . The above two types of representations are jointly input into the MLP prediction layer to output the sales volume prediction value:

$$\widehat{y}=\mathrm{MLP}(\left[X_{input},h_T\right])$$

(5)

where $\widehat{y}$ is the predicted sales result,$\mathrm{MLP}(\cdot )$ is the MLP network,

$X_{input}$ is the static input, and$h_T$ is the time-coded vector of trend variables.

This structure integrates the time-series evolution of advertising indicators and the static behavior of market response behavior, realizes information cross-learning in the AI model, and supports the model's ability to learn from the "trend change-driven sales conversion". "The fused feature space has the ability to learn from the global trend and the static behavior of market response behavior. The fused feature space has the ability of global expression, which can be used for subsequent analysis of the importance of key features and explanatory modeling.

3.4. Training Strategy and Parameter Optimization

The training strategy is centered on "improving the fitting accuracy of response to advertising market behavior". In the first stage, XGBoost is used to complete the high-dimensional static feature fitting, and the RMSE objective function is used to adjust the parameters such as the maximum depth$D$ , the sub-sample rate$\gamma$ , and the learning rate$\eta$ . In the second stage, the AdamW optimizer is used for the iterative training of the time-series modeling structure (LSTM and hybrid model). The loss function is consistent with the weighted mean square error defined in the previous section and does not need to be defined repeatedly.

In order to adapt to the volatility and trend mutation characteristics in advertising market data, an exponential decay strategy is introduced to dynamically adjust the learning rate:

$${\eta }_t={\eta }_0\cdot \exp (-\lambda t)$$

(6)

where ${\eta }_0$ is the initial learning rate,$\lambda$ is the decay coefficient, and$t$ is the current iteration round. Dropout mechanism is used to improve the robustness to advertisement noise samples, which is set to 0.3, and the EarlyStopping strategy controls the stability of iterations to prevent overfitting. During the training period, all model intermediate states, error changes and feature gradient information are recorded to provide basic data support for subsequent interpretability analysis and feature influence factor identification.

4. Experimental Design and Result Analysis

4.1. Experimental Setup and Evaluation Index

The experimental data uses advertising platform placement logs and sales records, the sliding window length is set to 7 days, and the prediction target is T+1 sales. The ratio of training set to test set is 8:2, the experimental platform is NVIDIA T4 GPU, the upper limit of deep model training epoch is 200, and the batch size is set to 128.The regression evaluation indexes include RMSE, MAE, and the coefficient of determination R2R^2R2.The key indexes are given in the following formulas:

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{(y_i-{\widehat{y}}_i)}^2,}}MAE={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n\left|y_i-{\widehat{y}}_i\right|,}$$

(7)

where $y_i$ is the real sales volume and${\widehat{y}}_i$ is the predicted value. This setting provides a unified experimental standard for subsequent model comparison and characterization.

4.2. Experimental Analysis of Different Model Comparisons

The three types of models, XGBoost, LSTM and LSTM-MLP, are evaluated uniformly, and it is found that the depth structure fits more stably under the effect of advertising trend variables, and LSTM-MLP outperforms XGBoost by about 12% in terms of RMSE indexes.XGBoost has a larger error in responding to high-click advertisements, and LSTM can adapt to temporal variations to a certain extent. Figure 2 shows the distribution of the residual discrete hotness of each model under different CTR intervals, which shows that XGBoost concentrates the error density in the high CTR segment, reflecting the lack of static model's ability to respond to advertising hotspots, indicating that the introduction of the trend modeling structure is more adaptive to the sales volume prediction.

4.3. Evaluation of the Influence of Advertising Features on Sales Volume

Using the SHAP method to interpret the sales forecasting model, the weights of CTR, CPI, MaterialAge and Platform features are significant.CTR has a positive correlation with sales, and MaterialAge has a significant negative impact on sales after more than 10 days, which indicates that aging of materials will weaken the conversion effect. Figure 3 shows the ranking and direction of the marginal impact of key advertising features, the larger the feature value the darker the color, the right side of the distribution of the dense area indicates that the model predicted value to improve significantly. The ad response category variables occupy the main driving position in the prediction, indicating that the sales prediction is dominated by the placement efficiency index.

4.4. Time-Series Impact Study of Trend Variables

Using the SHAP method to interpret the sales forecasting model, the weights of CTR, CPI, MaterialAge and Platform features are significant. CTR shows a strong positive correlation with sales, suggesting that maximizing exposure-to-click efficiency is a core target. MaterialAge has a negative impact after 10 days, indicating that outdated creatives reduce engagement, thus requiring timely material refresh. In practice, advertisers can dynamically allocate budgets toward high-CTR segments and prioritize platforms with better CPI performance. Figure 3 illustrates the marginal contribution ranking of each feature. The dominance of response-type variables suggests that optimizing click-through and conversion rates remains the most effective strategy for boosting returns.

5. Conclusion

Based on the in-depth modeling of the complex relationship between multi-source characteristics of the advertising market and game sales, a sales prediction framework that integrates the time-series characteristics of trend variables is constructed. Through structured mining and dynamic modeling of key indicators such as CTR, CPI, and material aging, the proposed multi-source fusion method effectively improves the accuracy and explanatory power of sales prediction. The experimental results show that there is significant heterogeneity in the performance of different features at the time lag level, and the trend variables possess higher stability in multi-step prediction. Although the model solves the problems of feature redundancy and time-lag mismatch to a certain extent, it is still limited by the fluctuation of data dimensions and the lack of generalization ability in cold-start scenarios. Future research can introduce cross-modal advertisement material content analysis and causal inference mechanisms to enhance the predictive model's ability and robustness to high-dimensional unstructured information.