← Portfolio · Plantation Analytics
Data analytics portfolio · Oil palm plantations, Indonesia

Turning plantation payroll and production records into decisions

A 2021–2024 analytics engagement with an oil palm plantation group: how efficiently harvesters work, why they leave, what fresh fruit bunch (FFB) output to expect next, and how to explain all of it to estate managers, HR and finance.

Internship project. Completed during my internship as a Trainee Data Analyst at Goodhope Asia Holding PTE LTD (February–June 2024). All worker data is aggregated or anonymised; no names or employee IDs appear on this site.

Python · pandasscikit-learn · K-Means · Random Foreststatsmodels · SARIMAXPlotly · treemaps · 3-D scatterPayroll / HR data

Four things this work demonstrates

Each card points to the section with the evidence.

Harvest efficiency

One estate group's harvesting workforce, month by month from January 2021 to March 2024, combining checkroll (payroll) records with weighed FFB output.

FFB harvested per month (MT)

Tonnes of fresh fruit bunches, whole estate group

Harvesters on payroll

Headcount paid in the month

Output per harvester (MT / month)

FFB tonnes ÷ headcount

Harvest pay per tonne (Rp thousand)

Piece-rate harvesting pay ÷ FFB tonnes; currency assumed IDR

Seasonality: when the bunches come

Average FFB by calendar month, 2012–2023, indexed so 100 = the estate's own monthly average

Agro BukitAgro Indomas

Domain notes behind the numbers

  • Piece rate vs assured amount. Harvesters are paid per bunch or tonne (activity amount) with a daily-wage floor (assured amount); the two streams must be separated before comparing productivity.
  • Cutting and carrying. Cutting and carrying labour is recorded per block per day, which is what makes block-level yield maps possible.
  • Crop cycle lag. Water stress is reflected in bunch output many months later, so a drought year distorts the following year's harvest and workload.
  • Working days and rainfall change what a month's headcount can deliver, so they were used as candidate drivers in forecasting.

Workers: who they are and why they leave

Recorded reason for leaving

Age at leaving

Months on payroll: leavers vs stayers

Share of each group, months paid within the Jan 2021–Mar 2024 window. Longer than 36 months is capped by the window.

LeftStill active

Where workers come from

Recruitment channel of leavers

Count of leavers by how they were recruited, split by main reason

Leavers by quarter

The attrition register covers Apr 2023 – Mar 2024 only, so quarters before that are not missing data, they are out of scope.

Worker segmentation and leaver-risk modelling

Four harvester segments (K-Means)

Each dot is a harvester: months on payroll vs total FFB harvested (log scale). Hollow ring = left the company. 700-worker random sample, no identifiers.

Segment profile

Choosing k: silhouette score

Higher = better separated clusters. k = 2 scores best; k = 4 was used because it gives segments a manager can act on.

Predicting who leaves: Random Forest

Features: months on payroll, total harvest pay, total FFB tonnes. Target: left vs still active. 25 % hold-out, stratified.

Default settings
Class-weighted (re-run)

Forecasting FFB production

Up to twelve years of monthly FFB output and rainfall for two estates. Seasonal ARIMA (SARIMAX) was used to project output, and tested the way it should be: on a year it had not seen, against a naive "same as last year" baseline.

FFB history and 36-month forecast

SARIMAX(1,1,1)(1,1,1)₁₂ fitted on 2012–2023. Shaded band is the 80 % prediction interval.

ObservedForecast

Back-test

ActualSARIMAXSame as last year

Does rainfall lead output?

Strength of correlation between monthly rainfall and FFB output k months later (k on the axis)

Annual FFB (thousand MT)

Annual rainfall (mm)

Working between developers and the business

The data came from payroll and HR systems built for accountants, and the audience was estate managers, HR and finance. Most of the work was in the gap between those two.

1 · Translate

Business terms → data model

Before any analysis I mapped the vocabulary used in the source files, so every column meant one thing to both sides:

BWBasic wage
SLSingle-line payslip
JVJournal voucher / entry
CKRCheckroll
DLSDaily log sheet (driver pay)
Assured amountDaily wage floor
Activity amountPiece-rate pay
2 · Structure

Mirror how the business reports

Analysis was split the way the organisation is split, so each stakeholder could check their own numbers:

  • Per estate unit and per month, then rolled up
  • Workforce, harvest and attrition kept as separate lenses
  • Block-level views for estate managers; group totals for finance
  • Cross-referenced with the company's own reference material (block maps, crop-cycle and El Niño reports)
3 · Deliver

The right format per audience

One analysis, several outputs:

  • Interactive HTML pages (treemaps, 3-D block plots, forecasts) that open in any browser, no software needed
  • A short slide deck on adjusted FFB production for business readers
  • Grouped CSV / Excel extracts for HR and payroll teams to keep working in familiar tools
Takeaway: the hardest requirement is rarely technical. It is agreeing what "productivity" or "left the company" means for payroll, HR and the estate, and writing that definition down before building charts.

Interactive exhibits

Original outputs from the engagement, kept as standalone pages. They load on demand. Block codes only, no worker identifiers.

Method, caveats and what I'd do differently

What was done

  • Cleaned and joined monthly payroll workbooks (basic wage, single-line payslips, journal vouchers) with employee, recruitment and attrition registers.
  • Built per-worker histories (months active, total harvest pay, total FFB, blocks worked) and group-level monthly series.
  • Fitted K-Means segments, a Random Forest leaver classifier, and SARIMAX forecasts with ADF stationarity tests and grid-searched orders.
  • For this page, the aggregates were recomputed from the source data. Original notebook results are quoted where they differ.

Honest limits

  • SARIMAX did not beat the naive baseline on the 2023 back-test for either estate. It captures seasonality but not year-to-year level shifts, such as the unusually dry 2023 at Agro Bukit.
  • Rainfall shows weak linear correlation with later output, so the "rainfall as driver" hypothesis is not supported by this data alone.
  • My original validation notebooks gave inconsistent errors (about 14 % MAPE in one case, near 100 % in others), which points to a pipeline fault, so I re-ran a clean back-test here.
  • The original leaver classifier reported 86 % accuracy but only 8 % recall on leavers; accuracy was misleading because most workers stay.
  • "Months worked" counts months seen in a 39-month window, not true tenure. Leavers are only recorded from Apr 2023.
  • Recruitment-channel counts have no baseline for stayers, so they show composition, not risk.

Next steps

  • Add rainfall and water-deficit lags with explicit 8–12 month terms and test on several held-out years.
  • Survival analysis (time-to-exit) in place of a binary leaver flag, with a proper tenure start date.
  • Add housing, family-status and shift data to examine the lifestyle factors behind absconding and returns home.

Data handling

  • Worker names and employee numbers are not on this site. Everything shown is aggregated or an anonymous sample.
  • Raw payroll workbooks (tens of GB across duplicated copies) are not published.
  • Currency is taken to be Indonesian rupiah from context; the source files do not label it.