Abstract:This paper contributes to the literature on parametric demand estimation by using deep learning to model consumer preferences. Traditional econometric methods often struggle with limited within-product price variation, a challenge addressed by the proposed neural network approach. The proposed method estimates the functional form of the demand and demonstrates higher performance in both simulations and empirical applications. Notably, under low price variation, the machine learning model outperforms econometric approaches, reducing the mean squared error of initial price parameter estimates by nearly threefold. In empirical setting, the ML model consistently predicts a negative relationship between demand and price in 100% of cases, whereas the econometric approach fails to do so in 20% of cases. The suggested model incorporates a wide range of product characteristics, as well as prices of other products and competitors.
What problem does this paper attempt to address?
This paper attempts to solve the problems of demand estimation and dynamic pricing in the retail environment, especially when product price changes are limited. Traditional econometric methods perform poorly in dealing with such low - price - change situations, and this paper proposes a deep - learning method based on neural networks to improve this problem.
### Specific Problems and Solutions
1. **Challenges of Demand Estimation under Low - Price Changes**:
- **Problem**: Traditional econometric models rely on significant price changes to accurately capture consumers' responses to prices. However, in practical applications, due to pricing policies or regulatory limitations, the price changes of some products are very small, which makes it difficult for the model to accurately estimate demand.
- **Solution**: The author proposes a neural - network - based method to estimate the form of the demand function by extracting potential patterns in product features and environmental variables. This method performs well in the case of low - price changes and can significantly reduce the mean - squared error (MSE) of the initial price parameter estimation.
2. **Improving the Accuracy of Demand Forecast**:
- **Problem**: In the case of low - price changes, traditional econometric methods cannot consistently predict the negative correlation between demand and price.
- **Solution**: The neural - network model can predict the negative correlation between demand and price in all cases, while in 20% of cases, traditional econometric methods fail to do so.
3. **Incorporating Multiple Product Features and Competitors' Prices**:
- **Problem**: Traditional methods usually only consider the price changes of a single product and ignore the influence of other product features and competitors' prices.
- **Solution**: The proposed neural - network model not only considers the price of the product itself but also incorporates a wide range of product features and the prices of other competitors, thus more comprehensively reflecting the changes in market demand.
### Research Contributions
- **Theoretical Contribution**: This paper extends the literature on parameterized demand estimation and shows the advantages of deep learning in dealing with low - price - change situations.
- **Empirical Contribution**: Through simulation and empirical analysis, it proves the superior performance of the machine - learning model in a low - price - change environment, especially in reducing the mean - squared error.
- **Application Contribution**: It provides retailers with more effective dynamic - pricing tools, especially in highly competitive market environments such as e - commerce.
### Summary of Mathematical Formulas
- **Demand Function**:
\[
D_i(P_i, X_i, S_i)
\]
where \( P_i \) represents the price of product \( i \), \( X_i \) represents environmental variables, and \( S_i \) represents product - specific features.
- **Neural - Network - Model Output**:
\[
\theta_i = FFC(F_{emb}(S_i), X_i)
\]
where \( F_{emb} \) is the embedding layer and \( FFC \) is the fully - connected layer.
- **Loss Function**:
\[
L(q_i, D_{\theta_i}(P_i)) = \frac{1}{n} \sum_{i = 0}^N (D_{\theta_i}(P_i) - q_i)^2
\]
where \( q_i \) represents the sales quantity of product \( i \).
- **Econometric Model**:
\[
\ln q_{jt} = \alpha_j + \beta_j \ln p_{jt} + f(\bar{d}_j, \gamma)^\top \bar{p}_t + \epsilon_{jt}
\]
where \( p_{jt} \) represents the price of product \( j \) at time \( t \), \( \bar{d}_j \) represents the distance vector, and \( \bar{p}_t \) represents the price vector of other products in the same category at time \( t \).
Through these methods, the paper successfully solves the problems of demand estimation and dynamic pricing in a low - price - change environment and provides more effective tools for the retail industry.