Data & Method
How the six baseline files become decisions, every data-quality fix applied, and every assumption — so each number can be traced and challenged.
Baseline data used
| File | Role in the brief | How the model uses it |
|---|---|---|
| BoM | Spare Part Master + Current Inventory Position | Stock, static Min/Max, MRP type, Installed Qty per FuncLoc (installed-base population) |
| FuncLoc list | Equipment criticality | A/B/C/S criticality (A highest) → part criticality, equipment class for failure-rate pooling |
| Notif History | Failure history | Failure rate per FuncLoc (installed-base prior), emergency frequency (downtime calibration) |
| Movement Material | Consumption history | Demand events, recency weighting, abnormal-issue detection, order type split |
| PR PO | Procurement & lead time | Unit price, lead time mean/variability, supplier reliability, open POs (on order) |
| Mitem & Mplan | Maintenance plan | Next due dates of part-consuming jobs → known future demand |
Planning date 01 Oct 2026; history 01 May 2025 – 30 Sept 2026. Database built 2026-11-04 14:10:20.
Pipeline
- 1. Clean & link
Deduplicate BoM and PR PO, normalize 28 cycle formats, convert currencies, link Notif → Order → Movement and BoM → FuncLoc.
- 2. Split demand by cause
Turnaround issues become event demand; preventive demand is replaced by the maintenance plan forecast; the rest is failure-driven.
- 3. Forecast irregular demand
Recency-weighted compound Poisson per part, blended by credibility with an installed-base prior (failure rate of every FuncLoc the part sits on). Never-used parts get a data-estimated first-demand rate.
- 4. Add known future demand
Part-consuming jobs from the maintenance plan, mapped to the parts they replace, at a usage factor calibrated on history.
- 5. ML early warning
Gradient boosting trained on 12 monthly snapshots (features before the date, label = demand in the next 90 days), validated out-of-time, then refit and scored on the planning date.
- 6. Optimize each part
Lead-time demand distribution (Poisson for slow movers, normal otherwise) → safety stock, reorder point, EOQ order quantity, Max. The service level is chosen to minimize holding + ordering + downtime-risk cost, or set by criticality.
- 7. Identical stock first
Material numbers with the same description and unit are linked; a need is covered by a twin's stock above its own reorder point before anything is bought.
- 8. Recommend & route
Reorder / transfer / reduce / monitor / maintain with reasons, health status, stockout and overstock risk; worklists per role, PR draft, decision impact, scenarios.
- 9. Prove by replay
The policy is rebuilt from data before 1 May 2026, then the actual May–Sep 2026 issues are replayed day by day against today's Min–Max and the Sparely policy.
Data-quality checks and fixes
| Dataset | Check | Result | Action |
|---|---|---|---|
| FuncLoc list | Missing criticality | 33 | Set to S (lowest) |
| BoM | Duplicate FuncLoc x Material rows | 230 | Removed (identical values) |
| BoM | Stock/Min/Max identical on every FuncLoc row of a material | True | Treated as warehouse-level values; counted once per material |
| BoM | Non Stock materials (min=max=0) still holding stock | 1683 | Flagged as uncontrolled stock in Inventory Health |
| Movement Material | Exact duplicate rows (same date, part, qty, order) | 863 | Removed (2.3% of issued units) |
| Movement Material | Movement types present | Issue | All rows treated as consumption (goods issue) |
| Notif History | Notifications without an order | 553 | Kept; no part link |
| Notif History | Notifications in May 2025 (backlog spike, ~4x normal) | 1463 | Failure rates measured from 2025-06-01 |
| Mitem & Mplan | Distinct cycle text formats | 28 | Normalized to days |
| Mitem & Mplan | Unparseable cycles | 0 | Excluded from schedule |
| PR PO | Exact duplicate rows | 5420 | Removed; 6,189 unique lines kept |
| PR PO | Non-IDR lines | 51 | Converted to IDR with fixed assumed rates |
| PR PO | PO delivery before PR date / lead time over 365 days | 2 / 106 | Kept as records, excluded from lead-time statistics |
| PR PO | Requested delivery date before PR date | 324 | Kept; lead time uses PO delivery date |
| PR PO | Open POs (delivery on/after as-of date) | 766 | Counted as on-order quantity in the inventory position |
| PR PO | Suppliers per material (max) | 1 | Single-source risk flagged |
| Movement Material | Issue lines linked to a notification via Order | 16.0% | Equipment link for the rest comes from BoM FuncLoc |
| Mitem & Mplan | ABC Indicator equals FuncLoc criticality | True | FuncLoc criticality used as the single criticality source |
| Mitem & Mplan | Planned-task part usage factor (actual preventive issues / installed qty of due tasks) | raw 0.716, 91/601 materials overlap | Factor 0.716 applied (clipped 0.05-1.0) |
| Movement Material | Parts unused before the split that got a first demand after it | 267 of 5897 (split 2026-05-01) | Sets the demand prior for never-used parts |
| ML early warning | Out-of-time AUC / top-decile precision | 0.895 / 47.6% (base rate 7.1%) | Trained on 59,848 part-months, tested on 14,962 |
| Movement Material | Last turnaround window (turnaround-type issue lines > 3x monthly median) | 2025-11-01 to 2026-01-31: 360 parts, 145,095 units; 36% of all units issued in 17 months fell in this window | Used as the scope for the next-turnaround readiness check |
| Replay | Policy inputs built from data before the replay window | 2026-05-01; 3,035 part-days of actual demand to replay | Both policies replayed day by day on the same demand and lead times |
| BoM | Material numbers sharing an identical description and unit | 133 groups, 597 materials; 66 groups have one number at zero stock while another holds stock | Identical stock is used before buying; groups flagged for master-data review |
| PR PO | Unit price of short-notice PRs (requested <= 14 days) vs the part's median price | ratio 1.000 over 1,021 lines | No rush premium in the data: default emergency premium 0% |
Assumptions
| Area | Assumption | Value | Note |
|---|---|---|---|
| Planning | As-of date | 2026-10-01 | Data ends 30 Sep 2026 |
| Planning | Forecast horizon | 365 days | |
| Demand | Order type by order-number prefix | 201=Corrective; 202=Emergency; 203=Turnaround; 204=Preventive; 205=Other maintenance; 206=Other maintenance; 209=Other maintenance | Not confirmed by organizers |
| Demand | Turnaround issues | Excluded from baseline, shown as event demand | |
| Demand | Recency half-life | 180 days | Recent demand weighs more |
| Demand | Installed-base prior calibration (kappa) | 0.4294 | Scales FuncLoc failure rates so total prior events match observed events |
| Demand | Credibility constant | 2.0 | Events needed before history outweighs prior |
| Demand | First-demand rate for never-used parts | 0.04092 per exposure-day | 267 of 5897 parts unused before 2026-05-01 were demanded after it |
| Demand | Failure-rate window | 2025-06-01 to 2026-09-30 | May 2025 backlog spike excluded |
| Maintenance plan | Part-consuming tasks | lube oil change -> OIL; re-greasing -> GREASE; filter element replacement -> FILTER ELEMENT, SCREEN; final element overhaul -> DIAPHRAGM, GASKET, O-RING, PACKING SET, SEAT RING, SPRING, STUDBOLT, TRIM SET; pump overhaul -> BEARING, COUPLING, GASKET, IMPELLER, O-RING, RING, SEAL, SLEEVE, STUDBOLT; (compressor|turbine) major overhaul -> BEARING, BLADE, DIAPHRAGM, FILTER ELEMENT, GASKET, NOZZLE, O-RING, PACKING SET, RING, SEAL, SLEEVE, STUDBOLT | Inspection/calibration/test consume none |
| Maintenance plan | Planned-task usage factor | 0.716 | Calibrated from preventive-order history |
| Procurement | PO-Item Amount | Unit price | Confirmed by organizers |
| Procurement | Lead time | PO delivery date - PR date | Blended with category when < 3 POs |
| Procurement | Expedited (emergency) delivery time | 10th percentile lead time of the part's category | Fastest deliveries actually observed; = downtime days when a stockout stops production |
| Procurement | Emergency price premium | 0% (short-notice price ratio 1.000) | Measured on 1,021 PO lines requested within 14 days |
| Procurement | FX rates to IDR | USD=16300; EUR=17700; SGD=12600; JPY=108 | |
| Procurement | PR type meaning | PR01=Import / long-lead; PR03=Regular replenishment; PR04=Regular (other); PR05=Project / turnaround; PR06=Regular replenishment; PR09=Urgent; PR10=Other | Not confirmed |
| Criticality | Ranking | A > B > C > S | Confirmed by the team; critical = A/B; a material takes the most critical FuncLoc it is installed on |
Safety and governance rules
Company policy, not cost–risk scenarios: fixed rules applied on top of the cost model.
| Rule | Definition | Today | Effect |
|---|---|---|---|
| Safety-critical element (SCE) | Installed on a FuncLoc tagged PRV- or SDV- (safety / pressure-relief valve, shutdown valve) | 860 parts on 1,040 locations; 578 of these locations are rated S | Downtime risk costed as class A, so 399 SCE parts are reorders today (Rp 8.9B); at least 1 spare always kept (1 part held only by this floor); never “potentially obsolete”; every reduction needs sign-off |
| Long-lead part | Critical (A/B or SCE), 80th-percentile lead time over 90 days measured on its own POs (category estimates do not count) | 544 parts | Badge and reason: the fastest delivery in its category is too slow for emergency buying, so plan ahead or hold a capital spare |
| Management sign-off | Buys above Rp 1.0B; reductions above Rp 1.0B or of any SCE part | 22 buys and 985 reductions among today's recommendations | Pending-approval list in the Management workspace; the PR draft and released capital wait for the sign-off |
| Decision log | Every approve, reject, undo and sign-off with role, time, quantity, value and reason | Kept in the browser in this MVP | Audit trail and CSV export in the Management workspace; a shared database in the pilot |
The SCE tag rule stands in for the plant's SCE register, which replaces it in the pilot. 56% of the tagged valves carry the lowest equipment criticality, so criticality alone would under-protect them.
Master-data work list
Candidates for the SAP master-data team to check, not automatic changes.
| Issue | Material numbers | Stock value | Suggested SAP action |
|---|---|---|---|
| Same description and unit under several numbers | 597 | Rp 38.9B | Check interchangeability; keep one active number and link or block the others |
| Non Stock part with repeated or planned demand | 265 | Rp 69.1B | Change MRP type to Stock Item and load the recommended Min / Max |
| Stock Item that the policy would buy on demand | 1,030 | Rp 1.54T | Change MRP type to Non Stock; run the stock down |
| Non Stock part holding stock | 1,582 | Rp 2.12T | Stock outside Min–Max control: return, transfer or reclassify |
133 groups share an identical description and unit; in 66 of them one number is empty while another holds stock (the Planner already transfers between them). The same description is not proof of the same specification, so each group needs a check before merging. Stock values use PO prices, or category medians for parts never bought.
Data gaps and the additional data we propose
Each gap limits a specific result; the proposal names the source and the business reason.
| Gap in the baseline data | Effect today | Proposed additional data |
|---|---|---|
| No part list per maintenance task (Mplan); 56% of plan items sit on a FuncLoc that has a BoM | Planned demand is inferred from the BoM with a calibrated usage factor | SAP PM task lists / component lists per maintenance plan |
| 83% of parts were never purchased in the data period | Their value uses category-median prices (labelled “estimated”) | Material master standard price / moving average price (SAP MM) |
| No production value per hour of downtime | Downtime cost per day is a scenario input; sensitivity shown above | Unit production plan × margin per tonne (e.g. ethylene) per equipment system |
| Unmet demand and historical stock levels are not recorded | Stockouts in the replay are simulated from the policies' own starting levels | SAP stock history (MB5B) and reservation / backorder records |
| Every part has a single supplier in PR PO (124 suppliers) | Single-source risk can be flagged but not re-sourced | Approved vendor list / source list (SAP ME01) per material |
| No link from equipment to its process system (only site-level nodes) | Results cannot be split by process unit | FuncLoc hierarchy (superior functional location) |
Sensitivity of the decisions to the assumptions that cannot come from data
Production loss per downtime day, holding-cost rate and cost per purchase order are business inputs. How much do the recommendations move if they are wrong?
| Case | Reorder | Reduce | Parts changing action | Avg inventory | Production stops / yr |
|---|---|---|---|---|---|
| Downtime cost × 0.5 | 1,099 | 3,614 | 335 (4.5%) | Rp 210.6B | 0.6 |
| Base case | 1,160 | 3,505 | – | Rp 260.0B | 0.31 |
| Downtime cost × 2 | 1,218 | 3,432 | 284 (3.8%) | Rp 303.4B | 0.2 |
| Holding cost 15% / yr | 1,172 | 3,471 | 123 (1.6%) | Rp 277.0B | 0.26 |
| Holding cost 30% / yr | 1,126 | 3,573 | 194 (2.6%) | Rp 231.3B | 0.47 |
| Order cost × 2 | 1,155 | 3,499 | 23 (0.3%) | Rp 261.6B | 0.32 |
Halving or doubling an assumption changes the action of only the share of parts shown; the direction of the recommendations is driven by the data (demand, lead time, price, criticality, observed emergencies).
Downtime calibration
Production loss per day is a business input (scenario page). How often a stockout actually stops production is learned from data: 132 Emergency notifications in 487 days = 98.9 per year, which today's static Min–Max must reproduce.
Suppliers (single-source exposure)
| Supplier | PO lines | Parts | Median LT | On time |
|---|---|---|---|---|
| RAJAWALI INSTRUMEN INTERNASIONAL PT | 474 | 59 | 24 d | 39% |
| CAHAYA TEKNIK PERSADA PT | 374 | 62 | 41 d | 36% |
| CAHAYA INDUSTRI SENTOSA PT | 290 | 24 | 33 d | 42% |
| SAMUDRA SEJAHTERA INTERNASIONAL PT | 275 | 30 | 31 d | 41% |
| CAHAYA LOGAM PRATAMA PT | 260 | 30 | 58 d | 41% |
| KARYA ENERGI SENTOSA PT | 232 | 100 | 78 d | 38% |
| GEMILANG LOGAM MAKMUR PT | 177 | 22 | 27 d | 41% |
| GEMILANG TEKNIK LESTARI PT | 153 | 6 | 32 d | 39% |
SQLite database
data/caliber.db — read by every page and API
| Table | Rows |
|---|---|
| assumption | 18 |
| backtest | 3 |
| consumption_by_type | 77 |
| consumption_monthly | 5,910 |
| data_quality | 25 |
| demand_event | 12,835 |
| forecast_monthly | 89,772 |
| funcloc | 14,484 |
| material | 7,481 |
| material_funcloc | 20,423 |
| material_twin | 3,614 |
| meta | 34 |
| ml_calibration | 7 |
| ml_features | 7,481 |
| ml_importance | 20 |
| ml_metrics | 3 |
| movement | 16,655 |
| mplan_item | 10,153 |
| notification | 5,795 |
| planned_demand | 826 |
| po_line | 6,189 |
| replay_demand | 3,035 |
| replay_material | 7,481 |
| replay_planned | 944 |
| supplier | 124 |
| tam_scope | 360 |
JSON API: /api/summary · /api/recommendations · /api/materials/[id] · /api/forecast · /api/replay · /api/scenario (POST) · /api/pr-draft (POST) · /api/impact (POST) · /api/data-quality · /api/master-data