Open research project · Uganda case study

Detecting fraud risk in public construction procurement, using open data

An open-source, explainable risk monitor that reads the public procurement record and helps auditors decide where to look. It flags suspicious tenders, suppliers and bidder groups, ranks contracts for audit, and points to ways procurement could move faster without weakening integrity.

237kopen-contracting records analysed (PPDA, 2015–2026)
5,400public works tenders · UGX 9.8 trillion awarded
18 + 2explainable red flags, plus 2 machine-learning detectors
119 / 135districts on the interactive map (4,096 contracts, about 70% of works value)

What the data shows

These are statistical risk indicators that call for review. They are not findings of wrongdoing against any entity or firm, and no firm is named on this site.

5.3×

Awards bunch just below the official estimate

1,976 works contracts were awarded at 97–100% of the entity's estimate, where a smooth distribution predicts about 316. The excess holds under open competitive bidding (4.5×) and in all 24 robustness specifications we ran (3.4× to 12×). Possible causes are leaked estimates, estimates revised after award, or bidders anchoring on a published budget.

Histogram of award-to-estimate ratios with a sharp spike just below 1.0 and a smooth counterfactual curve

205 rotating bidder groups

Identical groups of firms that bid together in 3+ procurements while the win passes between members. The largest group of three firms bid together 255 times for a single entity.

21 possible contract splits

Near-identical contracts from the same entity, awarded within 60 days without open bidding or to the same supplier. Found by text similarity (NLP) across 13 entities.

⅓ duplicate supplier IDs

24,707 supplier IDs collapse to 16,561 firms once spelling variants are merged. That has to be fixed before suppliers can be monitored reliably.

Visual stories

Three findings, animated. Each comes from a different branch of mathematics.

Statistics & calculus. Winning prices pile up just below the official estimate: 1,976 contracts where about 316 are expected.
Graph theory. The construction bidding network grows year by year; orange firms belong to groups that rotate wins. Firms are not named.
Data forensics. Procurement takes about 46 days; the award then waits a median of 101 days before it appears on the public portal.
Bar chart of what makes contracts unusual according to the Isolation Forest
Machine learning. What the Isolation Forest finds unusual: mainly suppliers that dominate an entity's works and long-standing supplier–entity ties.

Going further: five deeper questions

Beyond the red flags, we asked what the patterns cost, why they arise, whether Uganda is unusual, and whether the works were actually built. All results are risk indicators that call for review, and no firm is named.

Scatter of unit costs per latrine stance and per classroom with outliers highlighted
What should it cost? (NLP + statistics) Quantities extracted from contract titles give benchmarks: a latrine costs about UGX 5.2m per stance, a classroom about UGX 42m. 47 contracts cost more than twice the benchmark. Bunched contracts are not more expensive per unit.
Bar chart comparing the share of awards at 99–100% of the estimate across countries
Is Uganda unusual? (cross-country) The same screen on other countries' open data: Uganda's share of awards just below the estimate (23%) is about 3–4 times Rwanda's and Zambia's. Bunching is not universal.
Simulated distributions of winning prices with a known versus a secret estimate
Why does it happen? (game theory) In a model of bidding where firms know the estimate, prices pile up just below it. Keeping estimates secret removes the spike but makes some tenders fail, so secrecy is a trade-off. Estimated from real bids below.
Monthly shares of direct procurement and bunched awards around elections
Elections and the year-end rush (econometrics) In the six months before elections, awards exactly equal to the estimate rise by about 2 points. In April–June, very short bidding windows rise by about 4 points. Effects are modest; the charts show raw shares, the regressions add fixed effects.
Small multiples of road redness over time from Sentinel-2 for several road contracts
Was the road built? (satellite remote sensing) Copernicus Sentinel-2 imagery shows red gravel turning into dark asphalt. Hoima–Wanseko and Bumbobi–Lwakhakha were paved by 2021–22. Koboko–Yumbe–Moyo and Namagumba–Budadiri still look like gravel years after the tender; news reports show their works only started around 2024 (Koboko is due in July 2027), so the delay is mostly before construction.
True-colour Sentinel-2 crops of nine roads before, during and after paving or reconstruction
See for yourself (satellite images). One kilometre of each road, 1.2 km across, before, during and after. Hoima–Wanseko, Bumbobi–Lwakhakha and Muyembe–Nakapiripirit turn from orange gravel to grey asphalt; Luwero–Butalangu is being sealed now; Koboko–Yumbe–Moyo and Kalule–Bamunanika are unchanged; on Masaka–Mutukula the old asphalt has been stripped for reconstruction.
Stacked bars of asphalt, unpaved and vegetated shares along each road at three dates
Three-date semantic change (machine learning + remote sensing). A classifier trained on OpenStreetMap road surfaces (87% accurate on roads it never saw) labels each point along a contracted road as asphalt, unpaved or vegetated, before, during and after the contract. It agrees with the redness method: Muyembe–Nakapiripirit and Bumbobi–Lwakhakha are now mostly asphalt, while Koboko–Yumbe–Moyo and Namagumba–Budadiri are still unpaved. News reports show both started late (around 2024) rather than stopped, which is why the full tracking below checks every road against the media.
Map of Uganda with contracted road routes coloured by satellite status
Where the contracted roads are. 22 paving contracts and 5 reconstructions of paved roads, coloured by what the satellite shows in July 2026, after checking each road against news reports.
Heat strips: each row a kilometre of road, each column a Sentinel-2 reading from 2017 to 2026, coloured asphalt, unpaved or vegetation
Tender to tarmac: tracking construction kilometre by kilometre. For 22 contracted road sections, every kilometre is classified as asphalt (grey) or unpaved (orange) twice a year since 2017. Hoima–Wanseko was sealed in blocks in 2020–21 after earthworks in 2019; Luwero–Butalangu is being sealed from the Luwero end since 2025; Koboko–Yumbe–Moyo is unpaved along its whole length.
Timeline per road: tender date, works start, expected completion and satellite-observed completion
How long from tender to a sealed road? No road was complete within four years of its tender; the median is about 7.8 years. Where news reports give a launch date, works began a median 2.0 years (up to 5.5 years) after the tender: much of the delay happens before construction starts. Two roads show no paving after their expected completion and are candidates for inspection, not findings of wrongdoing.
Bars splitting each road's time into tender to works start and works start to completion
Where does the time go? Where a launch date is reported, works began a median 2.0 years after the tender, up to 5.5 years, before construction took another 2–5 years. Much of the delay happens before any machine is on site.
Two confusion matrices: blind visual check and OpenStreetMap tags against the satellite classifier
Is the classifier right? In a blind visual check of 160 random road crops it agreed 83% of the time where the image was clear, and it agreed with OpenStreetMap surface tags on 78% of 389 kilometres it never trained on. Many disagreements are recently paved roads that the map still tags as gravel.
Two scatter plots comparing media-reported progress and completion dates with satellite readings
Checked against the news. 39 dated reports on 17 roads. Satellite completion dates fall a median 0.6 years after reported completion or commissioning, and the sealed share tracks reported progress once paving starts (earlier works such as earthworks count in reported progress but are not sealing). The check also caught and fixed four wrong verdicts: a mis-located town, two roads finished before the 2017 archive, and a narrow road hidden by vegetation.
Satellite share of kilometres newly sealed against physical progress reported in government monitoring reports, by contract, and both over time
Checked against the government's own reports. Ministry of Finance monitors and the roads authority report physical progress on each upgrading project. Across 44 readings on 12 contracts, the satellite tracks them closely (rank correlation 0.84). Reported progress runs a little ahead, as it should: it counts earthworks and drainage before any asphalt.
For nine roads, the year each kilometre was sealed along the road, and a scatter of sealing pace against the longest pause
How roads get built, kilometre by kilometre. Once sealing starts, a road takes a median of three years at about 13 km a year, usually on several fronts at once. The long "pauses" turned out mostly to be roads let as two contracts, or overlapping a neighbouring contract. Measuring real stalls needs contract lot boundaries, which the open data do not publish.
Time to completion by dispute status; odds ratios of falling behind plan by the reasons monitors give; SHAP importance of a boosted model
What goes with slow delivery? (survival analysis + machine learning, exploratory) On 19 satellite-tracked roads, no procurement characteristic predicts completion time, but six completed roads is too few to tell. Across 249 monitoring visits to 167 projects, falling 10 points behind plan goes with design changes and contractor capacity problems. Land problems are cited everywhere, so they don't single out the worst projects. Monitors give reasons mainly when a project is behind, so these are associations, not causes.
For nine roads, a line from the launch ceremony to the first works visible in Sentinel-2, with arrows where none are visible yet
A ceremony is not a start. After each reported launch ceremony, when does the road corridor first change from space? Within one to three months on Masaka–Bukakata and Muyembe–Nakapiripirit; 2.8 years on Rwenkunye–Apac–Lira–Puranga; not yet, as of July 2026, on three roads launched in 2024–25. Earthworks on an existing gravel road can be hard to see, so these are upper bounds on the wait for any works.
Timeline of 23 road launch ceremonies with election dates, a histogram of months to the next election, and a placebo with shifted election dates
Launch ceremonies keep time with elections. Of 23 dated launch ceremonies of road works in 2018–2025, 11 (48%) fell in the 12 months before a general election, against about 20% expected if timing were unrelated (p = 0.003). Shifting the election dates lowers the share. Caveats: the ceremonies come from news reports, whose coverage may itself rise before elections, and a well-timed ceremony is not evidence that a road was delayed.
Map of Uganda's major roads with newly paved stretches coloured by whether they trace to a published contract, and a bar chart of the noise check
Every major road in Uganda: paving the record does not show. The same method run on all 15,573 km of trunk, primary and secondary roads finds 671 km of new paving between 2017 and 2026. Two-thirds traces to a contract in the open data; 230 km does not. Checked against the news, the longest untraced stretches are mostly real roads whose contracts are missing from the feed (the national roads authority appears as buyer in only 194 tenders in ten years), so the gap is in the record, not hidden roads. A noise check (apparent asphalt-to-gravel changes, which real roads almost never make) puts classifier noise at about a fifth of the signal.

Two more findings. If the bunching reflects prices anchored to the estimate, the excess cost is about UGX 10bn (95% CI 4.9–17.9bn) over 2015–2026. That's real money, but small next to total works spending; the bigger problem is information. And the Auditor General's engineering audits have mostly covered entities with lower risk scores than average. Audits follow donor programmes, not risk, which is exactly what risk-based selection would fix.

Road monitor: from tender to tarmac

Pick a contracted road. The map colours each kilometre by the year Sentinel-2 first read it as sealed (grey: never). The chart compares what the satellite sees with the physical progress the government's own monitors report.

A slow or paused road is a reason to ask for the work programme or visit the site, not a finding against anyone. The satellite sees sealing, not earthworks; reported progress includes earthworks, so it usually runs ahead. Disputes are listed by type from public court and tribunal records; no firm is named.

What if the estimate were secret? Evidence from 2,725 bids

Uganda publishes the estimated cost of what it buys. Using every bid in 790 auctions with two or more bids (all sectors, almost all from 2019, the one year with public bid-level data), we recover what bidders' costs must have been from how they bid (a structural auction model), then simulate other disclosure rules. Results are model-based and rest on five validity conditions tested in the paper; no bidder is named.

4.0×

Winning bids bunch just under the estimate, far more than losing bids

Just below the estimate there are about five times as many lowest bids as a smooth distribution predicts: an excess of 4.0 times the expected count (95% CI 3.1–5.2). Losing bids show an excess a tenth as large (0.4 times), and most of them (63%) sit above the estimate. Anchoring is concentrated in the bid that wins.

Distributions of lowest bids and other bids relative to the estimate, with a spike just below the estimate for lowest bids only

Competition is the cheapest lever

Implied markups fall from about 27% with two bidders to 6% with six or more. One more bidder per auction would cut expected prices by 12.2% with no failed tenders.

A binding ceiling beats secrecy

Making the published estimate a hard maximum gets most of secrecy's price gain (−8.7%) at a third of its cost in failed tenders: 5.8% against 15.6% for a secret ceiling with ±10% uncertainty.

Robust, with limits

The model fits auctions with three or more bidders; dropping two-bidder auctions shrinks all effects by about a third without changing their ranking. Markup levels depend on unobserved project differences.

Median markup by number of bidders, falling from about 27% to 6%
Markups and competition. Median implied markup by number of bidders, with robustness variants.
Price change and share of failed tenders under alternative disclosure rules
Alternative rules. Price change and failed tenders under a binding ceiling, a secret ceiling, a published range, and one more bidder.
Secret ceiling: price falls and failed tenders rise as uncertainty about the ceiling grows
The cost of secrecy. The more uncertain bidders are about a secret ceiling, the lower the price and the more tenders fail because every bid is above it.

Interactive map: where public works money goes, and how it is spent

The map places 4,096 public works contracts (about 70% of all works value) in Uganda's 135 districts. Contracts procured by district, municipal and city councils sit in their own district. Contracts procured by central bodies (UNRA, ministries, NWSC and others) are placed using the place names in their titles. Pick a question, a measure, a year and a sector, then click a district.

▶ Watch: how to use the map (40 seconds)

fewer than 10 contracts: interpret with care   no data

Click a district

Hover over or tap any district.
How to read the map, and its limits
  • Colours split districts into five equal-sized groups (quintiles) for the chosen measure, year and sector. Darker means more of the measure, and for the risk measures, more concerning.
  • Comparison: the panel compares each district with the national median, taken across districts with at least 10 contracts.
  • Locating central-government works: when a title names several districts, its value is split equally between them. Some multi-district road lots are placed at the station named in the title. About 2,150 contracts (around 30% of value) name no place, such as headquarters works, and are not mapped.
  • Days until award is public counts from bid closing to the date the award was recorded on the portal, so it mostly reflects reporting delay rather than how long the procurement took.
  • Single bid comes from bid records, some of which hold only the winning bid, so treat it as an upper bound.
  • Risk measures are indicators that call for review, not findings of wrongdoing. No firm is named.

Boundaries (2020) and population (2022 projection): Uganda Bureau of Statistics, via geoBoundaries and HDX (CC BY-IGO). Base map © OpenStreetMap contributors.

How the mathematics and machine learning are used

The project began with a concept note proposing a specific toolkit of mathematics, data science and machine learning. The table maps each proposed principle to what it does in the system, where it lives in the open-source code, and whether it is built yet.

Proposed principleWhat it does hereCodeStatus
Statistics & statistical inferenceBunching estimator with bootstrap confidence intervals detects the excess of awards just below the estimate; robust outlier tests (median absolute deviation) find abnormal contract values; 24 robustness and 4 placebo tests.analysis.py
red_flags.py
● Built
Benford's LawFirst-digit test of each entity's award values (mean absolute deviation vs. Benford); the 10% of entities that deviate most are flagged.red_flags.py● Built
Graph theory & network scienceCo-bidding network (553 construction firms, 1,333 links); Louvain community detection finds closed bidder clusters; repeated bidder groups and rotating winners.network.py● Built
Linear algebra: spectral analysisThe network is a weighted adjacency matrix. Its leading eigenvector (eigenvector centrality) ranks the most influential firms in each component.network.py● Built
Calculus & optimisationAudit-plan optimiser: a 0–1 knapsack that picks the contracts covering the most value at risk within a set number of auditor-days. A polynomial counterfactual (least squares) is used in the bunching estimator.app.py
analysis.py
● Built
Machine learning: unsupervisedAn Isolation Forest learns what a normal works contract looks like across 10 features. Its top 10% overlaps the rule-based top 10% by 32% (3× chance; ρ = 0.31), and it surfaces 298 further anomalies, driven mainly by supplier dominance and long-standing supplier–entity ties.ml_models.py● Built
Natural language processing(1) A title classifier identifies construction works, since the data has no category field. (2) TF-IDF text similarity finds near-identical contracts split across awards. Next: extracting audit findings from Auditor General reports.classify.py
ml_models.py
● Built / ◐ next
Machine learning: supervisedA model estimating the probability that a contract or supplier is fraudulent. It needs labelled cases (Auditor General and Inspectorate of Government findings). The 43 PPDA suspensions are too few, and mostly document forgery.—◐ Planned
Probability & measure theoryRisk scores and the audit model treat risk as a probability proxy; bootstrap resampling gives the sampling distribution of the bunching estimate. A calibrated probability needs the supervised model.analysis.py◐ Partly
Number theory & finite fieldsBlind bid evaluation: bidders get anonymous codes via modular exponentiation in a 2048-bit prime field. A private conflict-of-interest check lets two offices find overlapping firms without revealing their lists (commutative encryption). Try it below.blind_eval.py● Built (prototype)
Game theory (auction models)A first-price procurement auction with the estimate as reserve price reproduces the bunching and simulates what secret estimates would save and cost.theory_sim.py● Built
Remote sensingCopernicus Sentinel-2 imagery, twice a year since 2017, dates the sealing of every kilometre of 22 contracted roads, times launch ceremonies against visible works, and runs a census of paving on all 15,573 km of major roads. Validated against blind visual checks, OpenStreetMap tags and dated news reports.road_progress.py
launch_timing.py
paving_census.py
● Built
Structural econometrics (auctions)Bidders' costs are recovered from the distribution of their bids (first-price auction inversion), then equilibrium bids are re-solved under alternative disclosure rules, with bootstrap confidence intervals.structural_bids.py● Built
Econometrics (fixed effects)Election, fiscal-year-end and COVID windows estimated with entity and year fixed effects and clustered errors.natural_experiments.py● Built
CombinatoricsEnumerating bidder-pair combinations to build the network and identify identical bidder sets.network.py
red_flags.py
● Built

Try it: blind bid evaluation (number theory in action)

Bid evaluation should judge bids, not bidders. Enter some bidder names and a tender ID. Your browser hashes each name into the group of quadratic residues modulo a 2048-bit safe prime p (RFC 3526), raises it to a secret key k (y = h(name)k mod p), and turns the result into a code name. The evaluation committee would see only the code names. The same firm gets a different code on every tender, so it can't be tracked across tenders, and without k the code reveals nothing about the name.

Runs entirely in your browser; nothing is sent anywhere. A fresh secret key is drawn each time you load the page.

Codes will appear here.

Could this shorten procurement, and make it more transparent?

3 months

before a works award appears on the public portal

Using PPDA's own process dates, a typical works procurement takes about 6–7 weeks from invitation to award (median 46 days; 28 of them after bids close). But the award only appears on the public portal a median of 101 days after it is made, and 82% of works tenders are entered on the portal only after the contract has been awarded. The public record arrives too late for anyone to act on it while it matters.

What would help

Real-time publication. If each stage were published as it happens (invitation, bid opening, evaluation, award), red flags could be raised before contracts are signed, not months later.

Risk-based processing. Low-risk procurements (no red flags) already move about as fast as medium-risk ones (31 vs 25 days after bids close); high-risk ones take longer (about 52 days). A fast lane with automated document checks for the clean majority, and closer scrutiny for the high-risk few, would focus effort where it matters.

Correction (10 Nov 2026). An earlier version of this page said a works procurement takes a median of 206 days. That figure used the date an award was recorded on the portal, not the date it was made. Checked against PPDA's own process dates (a random sample of about 5,000 tender records), the procurement itself takes about 6–7 weeks; the remaining months are reporting delay. The figures above are corrected.

The product

Audit priority list

Every works contract ranked by risk weighted by value, with the reasons it was flagged.

Audit plan optimiser

Picks the contracts that cover the most value at risk within a set number of auditor-days.

Collusion signals

Repeated bidder groups, rotating winners, co-bidding networks and name-variant bidders.

Entity & supplier profiles

Concentration, method use, single-bid rates and Benford's Law checks.

An honest limitation

We tested the flags against providers that PPDA has suspended, and they did not score higher. Most of those suspensions (36 of 43) were for forged documents, which price and competition screens can't see. A complete system needs automated document verification alongside statistical screens.

Research outputs (in preparation)

Four working papers, each standing alone:
Bunching at the Estimate: Evidence, Mechanisms and Costs of Price Anchoring in Uganda's Public Works Procurement
Seeing Too Late: Transparency Lags, Risk-Based Oversight and Verification in Uganda's Public Works
Can Satellites Audit Public Works? Measuring Road Delivery and Delay from Open Contracting Data in Uganda
What Does Publishing the Estimate Cost? Bids, Markups and Disclosure Rules in Uganda's Procurement Auctions

About

Gava Charles: project concept, research design and direction. The project grew out of the concept note proposing a mathematical and machine-learning toolkit (number theory, linear algebra, graph theory, optimisation, statistics and ML) to help oversight institutions minimise losses from procurement fraud. It starts with Uganda's construction sector as the case study.

To cite: Gava Charles (2026). Procurement Risk Monitor: Uganda public works case study [software and data analysis]. https://github.com/gavacharles/procurement-risk-uganda