Stochastic Lot-Sizing Models

• Stochastic lot-sizing models are optimization frameworks that schedule order timings and sizes under random demand by balancing setup, holding, and shortage costs.
• They leverage dynamic programming, MILP, and scenario-based optimization to address high-dimensional state spaces and pseudo-polynomial complexities.
• Empirical results indicate that static-dynamic policies with rolling-horizon re-planning yield near-optimal performance in non-stationary and volatile demand environments.

Stochastic lot-sizing models are the cornerstone of quantitative inventory control under demand uncertainty, providing the algorithmic and structural foundation for production and inventory planning in complex, time-dependent environments. These models address the optimal timing and sizing of replenishments over a finite planning horizon where period-by-period demands are modeled as random variables, with the aim of minimizing expected total costs that include setup, holding, and shortage penalties. Research over the past decades has produced a rich taxonomy of stochastic lot-sizing strategies, efficient solution methods integrating dynamic programming, mixed-integer linear programming, and scenario-based optimization, and robust experimental validation of these approaches across a wide variety of demand patterns and operational regimes.

1. Formal Problem Definition and Structural Results

The classical single-item, single-location stochastic lot-sizing problem considers a finite planning horizon of $N$ discrete periods ($t=1,\ldots,N$), with each period's demand $D_t$ drawn independently from a known distribution $f_t(\cdot)$. At each decision epoch, the planner may issue an order ($z_t\in\{0,1\}$) of size $q_t\ge0$, incurring a fixed setup cost $K$ and a per-unit cost $c$, though $c$ is often omitted in comparative studies to focus on policy structure (Dural-Selcuk et al., 2016). The system accumulates holding costs $h$ per unit carried over, and backorder or penalty costs $t=1,\ldots,N$0 for unmet demand, yielding the canonical expected-cost minimization objective: $t=1,\ldots,N$1 where $t=1,\ldots,N$2 is the end-of-period inventory (positive if on-hand, negative if backordered).

Scarf's (1960) dynamic programming framework proves that, under convex costs and backordering, the state-dependent optimal control is governed by a pair of sequences $t=1,\ldots,N$3, where the $t=1,\ldots,N$4 policy reorders up to $t=1,\ldots,N$5 when the inventory position drops below $t=1,\ldots,N$6. The core DP recursion is: $t=1,\ldots,N$7 The computational challenge remains the pseudo-polynomial complexity in the inventory state space and the demand support (Dural-Selcuk et al., 2016), which motivates the development of efficient approximate and exact algorithms.

2. Classification of Policy Strategies

Stochastic lot-sizing research delineates several classes of replenishment policies, each corresponding to a particular information structure and planned recourse:

Policy Type	Order Epochs	Order Quantity Decision
Dynamic uncertainty	State-dependent	Adaptive ($t=1,\ldots,N$8 if $t=1,\ldots,N$9)
Static uncertainty	Fixed in advance	Fixed ($D_t$0 for pre-planned order periods)
Static-dynamic (R,S)	Fixed epochs	State-responsive ($D_t$1 at order epochs)

• Dynamic uncertainty: The canonical $D_t$2 policy leverages all realized inventory information. Policy parameters $D_t$3 are period-specific and computed using DP or MILP (Xiang et al., 2017, Visentin et al., 2022).
• Static uncertainty: The $D_t$4 approach fixes all order periods and quantities at time zero. Classical algorithms use network-flow DP [Sox & Vargas] or closed-form approximations for Normal demands (Dural-Selcuk et al., 2016).
• Static-dynamic uncertainty: The $D_t$5 strategy (Bookbinder & Tan 1988) commits to review epochs in advance but adapts order-ups to observed inventory at those epochs, combining plan stability with state recourse. High-quality MILP models and piecewise-linearizations enable near-optimal performance under general demand (Rossi et al., 2013).

Rolling-horizon re-planning (i.e., periodically re-solving over the residual horizon using updated state information) can transform static or static-dynamic policies, sharply reducing optimality gaps by exploiting realized demand up to the current period (Dural-Selcuk et al., 2016).

3. Exact and Approximate Solution Methods

Modern research implements both exact and approximate algorithms, targeting the computational bottlenecks introduced by demand convolution and high-dimensional decision spaces.

• Dynamic programming (DP): Scarf’s recursion remains the gold standard for optimal $D_t$6 parameterization but is limited in practice by state space growth. State discretization [Bollapragada & Morton 1999] and tailored heuristics (e.g., Silver-Meal-type for myopic cycle cost minimization [Askin 1981]) offer scalable approximations, though classical dynamic heuristics fare poorly on highly non-stationary profiles.
• Mixed-integer programming (MILP/MINLP): Recent models reformulate the loss functions governing holding and backorder expectations as piecewise-linear constraints, efficiently solvable with commercial MIP solvers (Xiang et al., 2017, Rossi et al., 2013, Ma et al., 2020). Binary search augmentation allows scaling MILPs for large horizons by sequential parameter extraction (Xiang et al., 2017).
• Stochastic dynamic programming heuristics: The $D_t$7 computation is now tractable via K-convexity accelerations and memorization (Visentin et al., 2022), enabling exact parameterization for substantial horizon lengths.
• Graph-based algorithms: The conversion of $D_t$8 policy evaluation to a shortest-path problem on a DAG, augmented repetitively to resolve infeasibilities, produces exact or near-exact solutions with polynomial complexity and modest optimality loss in rare cases (Ma et al., 2024).
• Polyhedral and chance-constrained approaches: For settings with explicit service-level constraints, the static probabilistic lot-sizing problem admits a two-stage, recourse-based convex hull description with facet-defining inequalities that unify and strengthen previous classes of cutting planes (Liu et al., 2016, Zhang et al., 2019).

Many advanced models embed these algorithms within rolling-horizon or simulation-optimization frameworks to mirror the information-dynamic environments encountered in stochastic production control (Schlenkrich et al., 2024, Sereshti et al., 2022).

4. Empirical Performance and Managerial Implications

Large-scale numerical comparisons validate the relative effectiveness of competing strategies:

• Optimality gaps (relative to the DP baseline) for dynamic heuristics are $D_t$94–5%, while static-dynamic MILPs (Tarim–Kingsman, Rossi et al.) achieve average gaps as low as 1.6% and as low as 0.2–0.3% under rolling re-planning. Pure static (R,Q) policies are outperformed, with typical gaps of 13% or more (Dural-Selcuk et al., 2016).
• Re-planning transforms static/static-dynamic policies—by updating only the realized inventory and re-solving the model for the remaining horizon—bringing even static policies within 0.5% of optimal on average (Dural-Selcuk et al., 2016).
• Demand nonstationarity, higher variance ($f_t(\cdot)$0), and "lumpy"/"life-cycle" patterns increase the performance gap across methods, with static-dynamic policies displaying the highest robustness. Dynamic heuristics based on stationary demand assumptions (Bollapragada–Morton) underperform on non-stationary series.
• Hybrid, multi-location, and capital-constrained models: Policies integrating lateral transshipments (via two-stage DP/MILP with receding horizon) (Ma et al., 2022) and explicit capital flow/borrowing constraints (Chen et al., 2017) have been shown to maintain near-optimality with dynamic-uncertainty policies ($f_t(\cdot)$1), particularly when parameterized by simulation-optimization.

Empirical findings consistently indicate that static-dynamic policies combined with rolling-horizon re-planning offer the best compromise between solution quality, stability, and operational flexibility across a range of realistic demand scenarios.

5. Extensions: Multi-echelon, Multi-stage, and Service-Level Constraints

Stochastic lot-sizing models are extensible to multi-item, multi-stage (multi-echelon) systems and can accommodate a range of operational and service-level constraints:

• Scenario-based and multi-stage stochastic programming: Two-stage models with full recourse under demand scenario sets are widely adopted in rolling horizon environments, with receding-horizon SP/MILP strategies outperforming static MRP or deterministic approximations under volatile or "nervous" forecasts (Schlenkrich et al., 2024).
• Progressive hedging and scenario decomposition: Multi-stage problems with setup carry-over require decomposition algorithms (e.g., Rockafellar–Wets progressive hedging), with penalty tuning and scenario voting strategies closing to 1–2% of compact model costs on scenario trees up to several thousand leaves (Schlenkrich et al., 11 Mar 2025).
• Service-level and chance constraints: Joint service-level constraints (e.g., cycle-fill or non-stockout probability), product substitution, and chance-constrained formulations can be integrated via recourse-extended MILP or branch-and-cut, yielding reliable constraint satisfaction and significant cost savings, especially when substitution is selectively applied (Sereshti et al., 2022, Zhang et al., 2019, Liu et al., 2016).
• General demand distributions and robust models: Robustness to demand model mismatch is increasingly targeted, with multi-period newsvendor variants and generic loss-function linearization enabling tolerable performance degradation under mis-specified demand laws (Khokhlov, 12 Feb 2026).

6. Modeling Guidance and Practical Recommendations

Implementation of state-of-the-art stochastic lot-sizing solutions should consider:

• Policy choice: Static-dynamic $f_t(\cdot)$2 with periodic re-planning dominates across most settings, offering stability and low computational burden.
• Algorithmic parameterization: Use 4–8 segments for piecewise-linear loss-function approximations in MILP; scaling to large horizons ($f_t(\cdot)$3) may require horizon decomposition.
• Demand modeling: Non-normal and highly non-stationary cases are tractable with piecewise-linear MILPs provided empirical quantile estimates for convolution supports are available; scenario-based SP is preferred when rolling-horizon forecast updates are prominent (Schlenkrich et al., 2024).
• Solver tuning: For high-precision industrial applications, set MIP integrality tolerance to $f_t(\cdot)$4 and gap tolerance to $f_t(\cdot)$5; branch on epoch indicators first, then cycle-location binaries (Rossi et al., 2013).
• Capital and resource constraints: Incorporate explicit borrowing and cash-flow equations to avoid suboptimal order timing; tune transshipment or capacity settings to match observed demand volatility (Chen et al., 2017, Ma et al., 2022).

7. Impact and Current Research Directions

Stochastic lot-sizing is pivotal in operational research, production–inventory analytics, and the design of robust supply chains under uncertainty. Current research advances emphasize the tractability and tightness of MILP approximations for arbitrary demand and policy structures, practical rolling-horizon and simulation-optimization integration, and the explicit incorporation of financial, multi-echelon, and service-level constraints in dynamic production planning. The convergence of stochastic programming, polyhedral combinatorics, and advanced decomposition methods continues to expand the modeling frontier for industrial-scale, non-stationary, and highly volatile environments (Dural-Selcuk et al., 2016, Xiang et al., 2017, Ma et al., 2024).