Which purchase orders are likely to arrive late?

Seven ideas behind this model

1The question

Which open orders will miss their due date if nothing changes, and which of them can still be saved?

A status such as "Ontime" or "Late" printed after delivery explains the past. A planner needs the opposite: a ranked list of orders that are heading late, a reason, and a date by which action still works. The model on this page builds that list from only five dates per delivery: the supplier promise date, the goods receipt at the hub (GR Offsite), the shipment start, the receipt at site (GR Onsite), and today.

2The business flow, end to end

Every shipment passes the same four phases. Each phase has an owner and a KPI. The model measures lateness along this chain.

Inbound and outbound are mirror images

Typical causes of delay, by phase (general experience, not company data)

3KPIs along the flow

Each KPI answers one question about one phase. Together they show who owns a delay.

4Method

A rule-based forecast, in four steps, applied to every PO line and every receipt.

Step 1. Group the raw data

Raw data has one row per receipt (RTN). If the same PO line appears more than once with different RTNs, the line was delivered in parts. The model keeps both grains.

• RTN grain carries the logistics assessment, because each delivery has its own dates.
• PO line grain carries the in-full decision and the site assessment, because quantity is ordered per line.

Step 2. Decide the stage of each PO line

Only events dated on or before the as-of date count. Quantities from all receipts of the line are summed.

Step 3. Forecast the arrival

Step 4. Compare with the target, then apply the bands

The tolerance before shipment is used for the vendor and hub stages; the tolerance after shipment is used on the water. Both are variables per hub, separately for the site and the logistics assessment.

5Formulas

Dates and standards

Quantities per PO line (events on or before the as-of date)

Forecast, slack and status

Where the days go: exact split of a late arrival

For a received delivery, the days late against the Due Site Date split exactly into four parts, because the Lead Time is the sum of the standard consolidation and the remaining standard:

Worked example, computed live from the hub table

Probability layer (extension, section 10)

6Parameter setup

Fill in this section before reading the results. A generic example is loaded.

6.1 General

6.2 Hub parameter table

One row per route. Std consolidation is the number of days goods are normally held at the hub; it is part of the Lead Time. Tolerances are the escalation window in days. Set 0 for a direct route.

6.3 Deliveries (one row per RTN)

PO-level fields (hub, PO quantity, promise date, lead time, cancelled flag, as-of override) are read from the first row of each PO line. Keep them identical on all rows of the same line. Cancelled = X removes the whole line. Leaving the as-of override empty uses the general date.

6.4 Probability layer settings (illustrative)

7Results and what-if

Move the levers to see which orders change status, and why. All results recalculate instantly.

7.1 What-if levers

7.2 Headline

7.3 Which orders are likely to arrive late

Ranked by probability of arriving late, then by slack. The status columns come from the rule-based model; the probability column is the extension in section 10.

7.4 Timeline of open and closed lines

7.5 Status counts, base against what-if

7.6 Logistics assessment per delivery, and where the days go

Columns on the right split the days against the freight table into supplier, GR posting, hub dwell against the KPI, and sailing against the standard. Negative = better than standard. A dash means the phase has not happened yet.

7.7 KPI chain from the data (not affected by the levers)

7.8 Self-check against the Excel model

8Reading the results

9Literature map

The rule-based model is the transparent baseline that recent research starts from. This map shows where three papers extend it.

10How to perfect the model

Nine gaps, what closes each one, and whether this page already shows it.

Back-test routine for upgrade 7

Store one row per PO line per day with its status. When a line closes, compare its status at day T−14 with the outcome:

Recalibrate each tolerance when the hit rate drifts. A lower tolerance raises the share of Late labels and lowers the share of missed late lines; the right value balances false alarms against missed warnings for your business.

Suggested build order

1. Capture the 3PL physical receipt date and the supplier's confirmed date (small data change, large attribution gain).
2. Start the daily snapshot now; it takes months to fill and every later upgrade depends on it.
3. Estimate spreads per route from history and replace the placeholder spreads.
4. Add the planned departure date at the hub, then vessel ETA.
5. Rank open lines by risk, value and need date.

11Limits and sources

Sources

• Intelligent Early Warning System for Supplier Delays Using Dynamic IoT-Calibrated Probabilistic Modeling in Smart Engineer-to-Order Supply Chains (journal).
• Conformal Predictive Distributions for Order Fulfillment Time Forecasting (arXiv preprint).
• Uncertainty-Aware Delivery Delay Duration Prediction via Multi-Task Deep Learning (arXiv preprint).
• Forecast error measures (bias, MAD, tracking signal) and dependability metrics follow common supply chain management practice.