Interdependent demand in the two-period newsvendor problem

The newsvendor problem is a classical task in inventory management. The present paper considers a two-period newsvendor problem where demand of different periods is interdependent (not independent), and seeks to follow this approach to develop a two-period newsvendor problem with unsatisfied demand or unsold quantity. Concerning the complexity of solution of multiple integrals, the problem is assessed for only two periods. In the course of a numerical solution, the probability distribution function of demand pertaining to each period is assumed to be given (in the form of a bivariate normal distribution). The optimal solution is presented in the form of the initial inventory level that maximizes the expected profit. Finally, all model parameters are subjected to a sensitivity analysis. This model can be used in a number of applications, such as procurement of raw materials in projects (e.g., construction, bridge-building and molding) where demand of different periods is interdependent. Proposed model takes into account interdependent demand oughts to provide a better solution than a model based on independent demand.


Introduction. The newsvendor problem is a classical problem in Operational
Research and inventory management [25]. Over time, this problem has undergone a significant development through the addition of new assumptions and constraints. The classical form of this problem seeks to determine the order quantity for a single product for a single period to maximize the expected profit under a series of assumptions, based on which demands are probabilistic and products are subject to discount or deterioration [11].
The appeal and importance of the newsvendor problem have encouraged a number of researchers to develop comprehensive literature reviews describing the wideranging extensions and modifications applied to the basic form of this problem. Most notable among these works are those by [25] Qin et al. and Khouja [16]. According to the classification proposed by Khouja, extensions have been developed through incorporation of factors, such as different objectives and utility functions, different supplier pricing policies, different newsvendor pricing policies [7], different levels of information about demand, constrained multi-products, multi-echelon systems, multi-location models, and multi-period models.
2. Literature review. The classical newsvendor problem cannot investigate scenarios where products remain unsold or demand remains unsatisfied over several periods.
2.1. Multi-period newsvendor problem. Matsuyama [17] developed a model for a multi-period newsvendor problem to deal with these issues. He assumed that some portion of products that remain unsold at the end of a period can be stored for the next period. He stated that, to determine the ordering quantity, one must first determine the amounts of products that remain stored or demand that remains unsatisfied from the last period. In his model, distribution of demand in different periods was assumed to be independent. Mileff and Nehéz [18] assessed an effective management of supplier corporations in a multi-product multi-period newsvendor problem. They proposed a heuristic-based optimal solution to satisfy the global capacity constraint policy. Altintas et al. [2] studied the problem of designing an optimal all-unit quantity discount with a single price break from the perspective of a distributor supplying a multi-period newsvendor buyer. The aim of this model was to maximize the supplier's profit, i.e., the income gained from the retailers minus the distribution costs. Behret and Kahraman [4] developed a fuzzy model for a multi-period newsvendor problem with a preseason extension and an emphasis on innovative products.This model optimizes the order period and order quantity by minimizing the sum of all fuzzy expected costs. Zhang [32] studied a multi-product newsvendor problem with both supplier quantity discounts and budget constraints. Given the presence of price discounts, the problem was formulated as a mixed-integer nonlinear programming model, it was solved by a Lagrangian relaxation approach. Zhang and Du [30] studied a multi-product newsvendor system with production capacity, where production can also be outsourced with zero or nonzero lead time. They developed the structural properties and solution procedures for profit maximization. Zhang and Hua [31] investigated a portfolio approach to a multi-product newsvendor problem with budget constraints, too, where the procurement strategy for each newsvendor's product is designed as a contract portfolio. They then proposed an efficient solution procedure and compared three models with different procurement contracts: a fixed-price contract, an option contract, and a contract portfolio. In some situations, the procurement strategy used for purchase can include a budget constraint such as a fixed price or an option or portfolio contract. This scenario has been modeled by Huang et al. [13], who demonstrated the advantage of a portfolio model. Ding and Gao [10] developed a methodology for determining the optimal (σ, S) policy for multi-product newsvendor problems that feature elements of uncertainty. They used uncertainty theory to deal with human uncertainty, which was given by the demand distribution. They designed their model such that ordering products triggered not only a linear ordering cost, but also a fixed setup cost, which depended on whether a joint or individual order was placed. Hanasusanto et al. [12] developed a risk-averse multi-dimensional newsvendor model for multiple products with strongly correlated demand. This problem was tailored for demand with strong dependence on mostly unknown future fashion trends whose distribution (called multimodal distribution) is in the form of spatially separated clusters of probability mass. This work, based on the assumption that distributional ambiguity will be addressed by minimization of the worst-case risks of order portfolios for all distributions compatible with the assumed modality. The NP-hard complexity of this problem was proven. An efficient, accurate, and conservative numerical solution was developed with quadratic decision rules. It was also shown that a solution that disregards ambiguity or multimodality may be unstable and fail to exhibit adequate quality and robustness under stress tests. Alwan et al. [3] formulated a newsvendor problem with correlated demand. They compared the performance of a traditional approach with a dynamic forecast-based approach. The authors showed that a minimum mean squared error (MSE) forecast model a better cost savings performance than the traditional approach. The performance of this MSE-optimal approach and the traditional approach was also compared with the performance of widely used alternative forecasting methods, such as the moving average and exponential smoothing methods. This article reported that when using alternative forecasting methods, sometimes the traditional approach to the newsvendor problem superior better results, and it is better to disregard correlations and forecasting. Orders for different periods show some interperiod dependency, which has been addressed in the recent literature ( [3,12]). Table 1 summarizes the literature dedicated to the inventory control and newsvendor problems. The approach most suitable for order planning is the multi-period newsvendor problem formulated by Matsuyama [17], in which interdependence of demand for different periods is disregarded.
The present study's innovation focuses on an incorporation of interdependence into the two-period newsvendor problem approach. The resulting formulations are applicable to large-scale projects where a majority of the raw materials is of the same nature.
To the best of the authors' knowledge, and as shown in Table 1, all multi-period and multi-product models available in the literature have assumed independent demand between periods. In many projects or manufacturing processes, materials and resources need to be ordered for two or more periods; so demand seems to be correlated. Therefore, to achieve a higher level of conformity with real-world applications, the present paper omits the assumption of independent demand and proposes a general solution for a two-period newsvendor problem with interdependent demand.
Section 3 presents the notations and assumptions used for formulation of the model. Section 4 describes the proposed model and proves the optimal solution. Section 5 presents the numerical results obtained with known variables. Section 6 discloses the results of our sensitivity analysis performed on all parameters. Some proofs are given in the Appendix.
3. Problem description. The significant distinction between the present paper and the one by Matsuyama [17] is the assumption that demand of two periods is correlated. When demand of two periods is interdependent, an increase in the demand of one period leads to a decrease in the demand of the other. Therefore, the following notations and assumptions are defined to develop the problem model.  [32] 1 Independent All-Units Zhang and Du [30] 1 Independent Zhang and Hua [31] 1 Independent Huang et al. [13] 1 Independent Sana [27] 1 Independent Chen and Ho [9] 1 1 Independent Ray and Jenamani [26] 1 1 Independent Ding and Gao [10] 1 Independent Kamburowski [14] 1 Independent Kamburowski [14] 1 Independent Pal and Sana [22] 1 Independent Hanasusanto et al. [

3.1.
Assumptions. This problem is modeled based on the following assumptions: • The assumptions of the classic newsvendor model. • There is a single product.
• In each project, the demand for a period depends on the demand for the other period. • Demand for independent periods is an independent random variable.
• Surplus supplies of a period will be used in the next period.
• Out-of-stock supplies of a period will be procured in the next period or will be lost. • The only objective is to maximize the mean profit from the planned periods. Selling price (per unit) of commodity at period j, p j : Buying price (per unit) of commodity at period j, s j : Holding cost (per unit) of commodity during period j, π: Penalty cost (per unit) of commodity when the demand is not met, C j : Setup cost at period j, x j : Total amount of demand during period j, f (x j , x j+1 ): Joint probability density function for demand, j = 1, α: Ratio of the amount stocked to the amount unsold, 0 ≤ α ≤ 1, β: Ratio of the amount sold at the beginning of the next period to the amount of unsatisfied demand in the present period, 0 ≤ β ≤ 1, δ: Ratio of selling price in each period, 0 ≤ δ ≤ 1, δq j +(1−δ)q j+1 : Selling price for the portion of demand that is not satisfied during the period j but is sold at the beginning of period j + 1, j = 1,

L:
Minimum demand in each period, N : Maximum demand in each period, k j , k j , A, B: Helping variable in differentiation of h js ; Decision variables: l j : Inventory level at the beginning of period j; Objectives: h js Profit at period j in state s, H j : Expected profit at period j, H: Total expected profit. This model is based on the assumptions of multi-period newsvendor problem, but here the main distinctive assumption is that f (x j , x j+1 ) = n i=1 f j (x i ), which means that the demand distributions are interdependent. Another assumption of the model is the presence of two periods, which simplifies the model.
Note that the contribution of the present paper is the provision of a general solution based on the assumptions above, which is not unlikely in real-world applications. In the next section, the profit function of the model is formulated according to the assumption below: In columns 2 and 3 of Table 2, L ≤ x j ≤ N means that demand and initial inventory of each period is usually limited to lower and upper bounds (L, N ). Equations (1)-(2) are defined accordingly and are considered to be one of the model's assumptions. Therefore, the following assumption is proved: j = 1: When demand in each period has lower and upper bounds (L, N ), initial inventory in period l j and the next period will have lower and upper bounds as shown in Equations (3)-(4), respectively; finally, it is proved in Equation (5) Depending on the status of demand, multiple conditions can be assumed for the model formulation; these conditions are shown in Table 2.
The second column in Table 2 describes a state where demand x j is less than initial inventory l j , sale income equals q j x j , buying cost equals p j l j , quantity of unsold products equals (l j − x j ), quantity of stocked products equals α(l j − x j ), holding cost of unsold products equals s j α(l j − x j ), penalty for unsatisfied demand is zero, and order of next period equals l j+1 − α(l j − x j ).
The third column in Table 2 describes a state where demand x j is greater than inventory level l j , sale income equals q j l j , buying cost equals p j l j , quantities of unsold and stocked products and holding cost of unsold products are all zero, unsatisfied demand equals β x j − l j , penalty for unsatisfied demand equals π(x j − l j ), and order of next period equals l j+1 + β(x j − l j ).
In columns 2 and 3 of Table 2, L ≤ x j ≤ N , meaning that the demand of each period is usually limited to lower and upper bounds (L, N ), which is one of the model's assumptions. Figure 1 shows the step-by-step process of implementation of the conceptual model.  Table 2. Conceptual Model.

Description
Period j Status of demand

4.
Mathematical model for interdependent demand in a Two-Period newsvendor problem. The main objective of the formulation of this newsvendor problem is to construct the objective function with respect to the range of demand and initial inventory level. In the formulation provided by Matsuyama [17] for a twoperiod problem with uncorrelated demands, h js (α, β, δ, π, l 1 , l 2 , x 1 , x 2 ) is assumed to be the profit function for period j and α, β, δ, π are assumed to be as follows.
In the first period, only two states are possible. In one state, demand is lower than the initial inventory (x 1 ≤ l 1 ) and profit equals sales (determined by demand) minus inventory purchase costs (Equation (6)). In the second state, demand is greater than the initial inventory (x 1 > l 1 ) and profit equals sales minus purchase costs, lost sales and setup costs (Equation (7)) [17]: j= 1 and s= 1 : j= 1 and s = 2 : In the second period, four states are possible. In the first state, demand of the first period is lower than the second initial inventory (x 1 ≤ l 1 ) and demand of the second period is also lower than the inventory level (x 2 ≤ l 2 ). As shown by Equation (8), here, overall profit equals profit of the first period minus holding cost plus sales of second period (with respect to its demand) minus initial purchase cost minus cost of the percentage of products that are moved to this period, and also minus setup costs [17]: j = 2 and s = 1 : In the second state, demand of the first period is less than the initial inventory (x 1 ≤ l 1 ), but demand of the second period is greater than the second inventory level (x 2 > l 2 ). As shown by Equation (9), here, overall profit equals profit of first period minus holding cost plus sales of initial inventory minus inventory purchase cost minus cost of the percentage of products that are moved to this period minus shortage cost of the second period and setup costs [17]: j = 2 and s = 2 : In the third state, demand of the first period is greater than initial inventory (x 1 > l 1 ) and demand of the second period is lower than the inventory of the second period (x 2 ≤ l 2 ). As shown by Equation (10), here, overall profit equals profit of first period plus sales of leftovers of the first period (with mixed price) plus sales of the second period minus costs of purchase of the second inventory and costs of purchase of the percentage of products that are moved to this period minus setup costs [17]: j = 2 and s = 3 : In the fourth state, demand of the first period is greater than the initial inventory (x 1 > l 1 ) and demand of the second period is greater than the initial inventory (x 2 > l 2 ). As shown by Equation (11), here, overall profit equals profit of the first period plus sales of leftovers of the first period (with mixed price) plus sales of the initial inventory minus cost of purchase of the initial inventory plus cost of purchase of the percentage of products that are moved to this period minus shortage cost of the second period and setup costs [17]: j = 2 and s = 4 : Demands of periods are assumed to be interdependent, so in the next step, the expected profit function has a bivariate demand probability distribution function In Equation (12), H (l 1 , l 2 ; x 1 , x 2 ) is a function expressing the expected profit of the second period. This function is a dual integral on initial inventory of the first and second periods and can be obtained by multiplication of demand distribution function by profit functions of the four above-mentioned states. In the first part of Equation (12), the profit function Equation (8) bounded by x 1 ≤ l 1 , x 2 ≤ l 2 has been multiplied by f (x 1 , x 2 ). In the second part of Equation (12), the profit function Equation (9) bounded by x 1 ≤ l 1 , x 2 > l 2 has been multiplied by f (x 1 , x 2 ). In the third part of Equation (12), the profit function Equation (10), bounded by x 1 > l 1 , x 2 ≤ l 2 , has been multiplied by f (x 1 , x 2 ). Finally, in the fourth part of Equation (12), the profit function Equation (11) bounded by x 1 > l 1 , x 2 > l 2 has been multiplied by f (x 1 , x 2 ).
Next, to obtain the optimal solutions for initial inventory of the first (l * 1 ) and second (l * 2 ) periods, it is necessary to differentiate H (l 1 , l 2 ; x 1 , x 2 ) with respect to l 1 , l 2 and equate this then to zero: To obtain optimal solution l 1 , the derivative of H is found: Equation (12) has been changed into parametric form in Equation (15), so H j (l 1 , l 2 ; x 1 , x 2 ) has been replaced with its short form h 2j (l 1 , l 2 ; (16)). In Equation (17), with respect to l 1 , it has been differentiated, and then a change in variable ∂ ∂l1 h 2j (l 1 , l 2 ; x 1 , x 2 ) with k j shown in Equation (18) results in Equation (19). This change of variable allows the dual integral of l 1 to be easily differentiated (Equation (20)): Considering the presence of ∂ ∂l1 H j (l 1 , l 2 ; x 1 , x 2 ) in Equation (20), the dual integral has been differentiated. Then, in Equation (21), it has been replaced with k j f (x 1 , x 2 ). Although this equation contains the terms H 1 (l 1 , l 2 ; l 1 , x 2 )−H 3 (l 1 , l 2 ; l 1 , x 2 ) and H 2 (l 1 , l 2 ; l 1 , x 2 ) − H 4 (l 1 , l 2 ; l 1 , x 2 ), it will be proven that these terms are zero.
After proving Corollary 1, Equations (26)- (27), the results need to be inserted into Equation (21), revealing the (l 1 , l 2 ) in The process of differentiation continues until ∂ ∂l1 H (l 1 , l 2 ; x 1 , x 2 ) becomes (k 1 − k 2 − k 3 + k 4 ) F (l 1 , l 2 )+(k 2 − k 4 ) F (l 1 )+(k 3 − k 4 ) F (l 2 )+k 4 (Equation (28)): In this step, given as above k 1 , k 2 , k 3 , k 4 , we differentiate h 2j with respect to l 1 (Equations (29)-(32)). As a result, k 1 = k 2 and k 3 = k 4 ; then these parameters are inserted into Equation (28): In the following, to obtain the optimal solutions for the initial inventory of the first and second periods l * 1 , in Equation (35), we need to make ∂ ∂l1 H (l 1 , l 2 , x 1 , x 2 ) equal to zero. after moving the known variables to the right side, as Equation (35) shows, F (l * 1 ) equals k 4 /(k 4 − k 2 ): Finally, it has been proved that the cumulative demand function (CDF) of l * 1 equals the constant parameter of Equation (36), and to determine l * 1 in Equation (36), F should be inverted. However, the infinity term in the above equation and the approximation of l * 1 are approximately in the following through numerical computation: Similarly, F (l * 2 ) can be obtained from Equations (47)-(74) in Appendix 1, and the final results are shown in Equations (38)-(39). Although the values of l * 1 , l * 2 have been computed, to determine whether l * 1 and l * 2 are extrema or optimal solutions, a Hessian matrix needs to be constructed with respect to l * 1 and l * 2 ([1, 23]): After computing the determinant of the Hessian matrix, since the determinant becomes positive and the values of (k 3 − k 4 ) and (k 2 − k 4 ) become less than zero, l * 1 and l * 2 develop into optimal solutions. According to L ≤ x 1 ≤ N and L ≤ x 2 ≤ N from Equations (1)-(2), l * 1 and l * 2 need to be determined by a truncated distribution with correlated demands, negative and positive infinity are changed to L and N [28]: In Equation (43), the function f (x 1 , x 2 | L ≤ x 1 ≤ N, L ≤ x 2 ≤ N ) has been replaced by f (x 1 , x 2 ) with upper and lower bounds of L ≤ x 1 ≤ l * 1 and L ≤ x 1 ≤ N , L ≤ x 2 ≤ N and L ≤ x 2 ≤ l * 2 : After truncating the distribution and replacing f ( in Equations (44)-(45), l * 1 and l * 2 have been approximated through numerical computation. The method by which numerical computation has been utilized is shown in Appendix 3.
In the following, the parameters have been inserted into the formulas. Additionally, according to Appendix 2, the optimal value obtained through two steps of numerical computation is l * 1 = 220.99 and l * 2 = 240.88. After computing the optimal solution, l * 1 and l * 2 should be inserted into Equation (46), where the value of H can be obtained as the expected value: Table 3 and Figure 2 show that proposed model leads to a better profit (H) than Matsuyama [17] our instance problems. As a result, a model that takes into account interdependent demand oughts to provide a better solution than a model based on independent demand.  Figure 2. Chart of Differences between the proposed model and Matsuyama [17].
However, we would like to recall that our results in Table 3 are based a bivariate normal distribution and numerical calculation; these results may change when distribution and parameters are becoming varied. The subject of future research will be about the question why the results obtained by our this model are better than the outcomes of previous methods, from a rigorous mathematical perspective.
We underline that this model can be used in a number of further applications, such as procurement of raw materials in projects (e.g., construction, bridge-building and molding) where demand of different periods is interdependent ( [19,21]).
6. Sensitivity analysis. The formulation developed for the proposed problem is based on a number of assumptions that involve several parameters; therefore, the sensitivity of objective functions to variations in the parameters needs to be analyzed. Table 4 shows the results of sensitivity analysis obtained by determining the expected profit after making slight changes in seven parameters. The most important results are presented in Table 4, and proofs of all formulations are presented in Appendix 3. Table 4. Sensitivity analysis of the proposed model. Table 4 shows that for all q 1 , q 2 the value of H (l 1 , l 2 , x 1 , x 2 ) is ascending and for π, p 1 the value of H (l 1 , l 2 , x 1 , x 2 ) is descending. The value of other parameters such as β and δ depend on p 2 − s 1 , (δq 1 + (1 − δ) q 2 ) − p 2 and q 1 − q 2 . Moreover, H (l 1 , l 2 , x 1 , x 2 ) is ascending whenever it is positive, and is descending otherwise. However, the parameter p 2 depends on other parameters and its status is unknown. Proofs of these sensitivity analyses are presented in Appendix 3.
Tables 5-7 and Figures 3-5 show a sensitivity analysis on expected value (H) of the ratio α, β, δ that have proved based on differentiation in Table 4. Parameters vary between 20% to 100%, both increasing and decreasing of the parameters is considerd in Tables 5-7.  Table 5 and Figure 3 reveal that expected value (H) increases when α increases; this has been proved based on differentiation in Table 4. Table 6 and Figure 4 show that the expected value (H) increases when β increases; this was demonstrated through differentiation in Table 4.
By Table 7 and Figure 5, we learn that the expected value (H) increases when δ rises, this was proved via differentiation in Table 4. All the above examples disclose that our model is verified and meets the requirements a good matematical model.    7. Conclusion and outlook. The present paper sought to develop the formulation provided by Matsuyama [17] to study the two-period newsvendor problem with unsatisfied demand or unsold quantity and interdependent demand as a new approach to this problem. Due to the complexity of the solution of multiple integrals, the problem was assessed for only two periods. The aim was to calculate the optimal value of the initial inventory at each period; because of the emergence of double integrals in the distribution function, the initial inventory was approximated by numerical computation. In numerical results, bivariate normal distribution was used to obtain the optimal solution, defined as the initial inventory at each period. Formulation of the proposed problem was based on a series of assumptions involving multiple parameters, so it was necessary to determine the sensitivity of the objective functions with respect to variations in the parameters. Therefore, all the parameters underwent a subjected to sensitivity analysis. The proposed model takes into account interdependent demand oughts to provide a better solution than a model based on independent demand. it can be used in a number of applications, including procurement of raw materials and products required by projects in the demands of periods are interdependent. Future research could use this approach with multi-period (n-period) models, which can be predicted to be in the form of multiple integrals. Another approach would be to use a fuzzy demand distribution function, another multivariate distribution to achieve more conformity with real-world applications or considering budget constraints [20]. Another plausible approach would be to develop the association of competitive environments with cooperative and non-cooperative game theory in multi-newsvendor environments. A further suggestion is to develop this model by adding quantity discounts and nonzero lead times to motivate customers to buy perishable goods in single-or multi-product problems. To obtain the optimal solution l * 2 with a process similar to that followed in Equations (6)-(36), H (l 1 , l 2 , x 1 , x 2 ) should be differentiated with respect to l * 2 : In Equations (48)-(49), H (l 1 , l 2 , x 1 , x 2 ) is differentiated, and H j (l 1 , l 2 , (50).
Let be k j = ∂ ∂l 2 h 2j (l 1 , l 2 , x 1 , x 2 ) ; then, In this step, h 2j has been differentiated with respect to l 2 , and the results are k 1 , k 2 , k 3 , k 4 in Equations (66)-(71). As a result, k 1 = k 3 and k 2 = k 4 in Equation (71), then these parameters have been inserted into Equation (65): Finally, it has been proved that the cumulative demand function of l * 2 equals a constant parameter in Equation (74). For determining l * 2 in Equation (74), F should be inverted. However, the infinity term in the above equation and approximation of l * 1 have been computed through the numerical computations provided in Appendix 2.
Let be A = Here, l * 1 has been received through almost two steps of numerical computation, until F (l * 1 , N ) − F l * 1 , N < occurs in Equations (75)-(79). Moreover, after truncation of the normal distribution, l * 2 also needs to be obtained through numerical computation: Now l * 2 has also been got through almost two steps of numerical computation, until |F (N, l * 2 ) − F N, l * 2 | < occurs in Equations (80)-(85).