Policies¶

A policy is an ordering rule: given the inventory state, how much should be ordered? Stockcast ships the classic rules of inventory theory and lets you write your own with the same interface.

At a glance¶

Policy	Notation	When it orders	How much
`OrderUpToPolicy`	\((R, S)\)	every scheduled opportunity	\(\max(0, S - \mathit{IP})\)
`ReorderPointPolicy("sQ")`	\((s, Q)\)	when \(\mathit{IP} \le s\)	\(Q\)
`ReorderPointPolicy("sS")`	\((s, S)\)	when \(\mathit{IP} \le s\)	\(\max(0, S - \mathit{IP})\)
`ReorderPointPolicy("sS")`, `review_period=R`	\((R, s, S)\)	at scheduled opportunities when \(\mathit{IP} \le s\)	\(\max(0, S - \mathit{IP})\)
`SingleOrderPolicy`	newsvendor	once	up to a season target
Your own	–	your rule	your rule

Choosing a policy¶

flowchart TD
    A{"Can you order<br/>more than once?"} -- no --> N["SingleOrderPolicy<br/>(newsvendor)"]
    A -- yes --> B{"Does each order have<br/>a fixed size, e.g. a pallet?"}
    B -- yes --> SQ["ReorderPointPolicy (s, Q)"]
    B -- no --> C{"Is placing an order<br/>costly compared with holding stock?"}
    C -- "no, order whenever there is a chance" --> RS["OrderUpToPolicy (R, S)"]
    C -- "yes, only order when low" --> D{"Fixed ordering days?"}
    D -- yes --> RSS["ReorderPointPolicy (R, s, S)<br/>review_period=R"]
    D -- "no, check every period" --> SS["ReorderPointPolicy (s, S)<br/>review_period=1"]

These are starting points. Because every policy runs through the same engine and event table, the fastest way to choose is often to simulate the candidates side by side with run_comparison.

The shared interface¶

Every policy is configured with the same operational facts:

Argument	Meaning
`lead_time`	\(L \ge 0\), periods from order to delivery
`review_period` or `schedule`	when the policy may order (see Decision schedules)
`service_level`	the probability \(\alpha\) its targets represent, or `None` for planner targets
`allow_backorders`	`True` for backorders, `False` for lost sales

Then it follows the same lifecycle:

policy = PolicyClass(lead_time=..., review_period=..., ...)   # configure
policy.fit(targets, ...)                                        # bind forecast information
decision = policy.predict(state, current_period=...)            # propose orders

predict returns an OrderDecision: one row per SKU with order_quantity and the reasoning behind it (target_level, inventory_position, reorder_point, order_period, expected_delivery_period).

Fitted policies also let you inspect what they learned:

Method	Returns
`get_target_levels()`	\(S\) per SKU (order-up-to and single-order)
`get_parameters()`	\(s\) with \(Q\) or \(S\) per SKU (reorder-point and periodic review)
`get_target_metadata()`	probability, horizon, origin, end date, source

Policies and stock¶

Policies decide; they never change stock. They receive a copy of the state and return a request. Constraints, callbacks, suppliers, receipts, and demand are all applied by the engine, after the policy. As a result, every policy is judged by the same accounting.

See also: Walkthrough step 4 · Notebook 05b · Notebook 06