Dataset

InventoryBench consists of 1,320 instances split between synthetic trajectories and real demand trajectories. Each instance defines a single multi‑period inventory game.

Instance families

Synthetic trajectories (720 instances)

Synthetically generated demand trajectories covering 10 distinct demand patterns, each with 4 parameter variants.

• 10 patterns × 4 variants = 40 distributions
• 2 demand realizations × 3 profit‑to‑holding‑cost ratios = 6 instances per variant
• 3 lead‑time settings
• Total: 40 × 6 × 3 = 720 instances

Real trajectories (600 instances)

Weekly sales data from H&M retail, covering 200 distinct product articles.

• 200 articles
• 3 lead‑time settings
• Random profit‑to‑holding‑cost ratios
• Total: 200 × 3 = 600 instances

Directory layout

The benchmark data is included in the repository under benchmark/:

benchmark/
├── synthetic_trajectory/
│   ├── lead_time_0/
│   ├── lead_time_4/
│   └── lead_time_stochastic/
└── real_trajectory/
    ├── lead_time_0/
    ├── lead_time_4/
    └── lead_time_stochastic/

Each leaf directory contains multiple instance folders. For example:

synthetic_trajectory/lead_time_0/p01_stationary_iid/v1_normal_100_25/r1_low/
  train.csv
  test.csv

real_trajectory/lead_time_0/108775044/
  train.csv
  test.csv

File formats

Each instance directory contains:

File	Description
train.csv	Historical demand samples (context for algorithms)
test.csv	Demand trajectory for evaluation

train.csv

exact_dates_{item_id},demand_{item_id}
Period_1,108
Period_2,124
...

• Synthetic: five samples from the underlying distribution.
• Real: five weekly observations preceding the test period (actual dates in YYYY-MM-DD format).

test.csv

exact_dates_{item_id},demand_{item_id},lead_time_{item_id},profit_{item_id},holding_cost_{item_id}
Period_1,108,0,4,1
Period_2,124,0,4,1
...

Columns:

• exact_dates: period identifier (synthetic) or actual date (real).
• demand: realized demand.
• lead_time: order lead time (e.g., 0, 4, or stochastic values).
• profit: profit per unit sold.
• holding_cost: holding cost per unit per period.

Lead times

We provide three lead‑time configurations:

• lead_time_0: immediate delivery (no delay).
• lead_time_4: fixed four‑period delay.
• lead_time_stochastic: per‑order lead times drawn from a small discrete set, including values that represent never‑arriving orders.

The same stochastic lead‑time sequence is shared across instances within each family to support fair comparisons.

Cost structures

Three profit‑to‑holding‑cost ratios are used:

• Low: 1:1 (balanced stockout vs holding costs).
• Medium: 4:1 (moderate stockout penalty).
• High: 19:1 (high stockout penalty for service‑critical items).

Synthetic instances cycle systematically through these ratios; real instances sample them randomly.

Using the dataset

• Clone the repository to get the full dataset under benchmark/.
• Use the directory and file naming scheme to iterate over instances.
• See the repository README for reference implementations and how to run benchmarks.