Nicolás Pavón

It's all about money

In today's world, corruption is an issue that affects every country. Therefore, having a standard to store government purchases, like the Open Contracting Data Standard (OCDS)., plays a vital role in uncovering fraudulent activities and financial misuse by the government. Luckily, in my home country, Uruguay, this standard has been used since 2015 to store a wide range of government purchases, and this data is publicly available for free use on the Datos Abiertos.

In this case study, we will analyze this data in search of interesting patterns and characteristics, which hopefully (or not), will reveal some misuse of government funds.

inspiration

While looking for a interesting dataset to demonstrate my ML skills, i came across the website Cuentas claras . In there, the owners try to reveal fraudulent purchases and misuse of government founds by applying a "function", which analyzes certain parameters and returns a value, in their case, "red flags". This idea of "red flags" is not new, and as we can see in the >OCDS website, this was already attempted by a few countries, and there are guides to work the OCDS data looking for these red flags.

Cuentas claras website

I strongly reccomend taking a look into the website and explore it. As you can see in the image above, they have succesfully processed and classified purchases of many public government organizations, marking them with the famous "red flags". Also there are some interesting statistics to explore.

Sadly the "Cuentas claras" proyect does not provide the processed data along with the "red flags", and after some failed attempts in contacting the owners, it seems to be abandoned or in standby state.

This is not good news. In order to train a predicting model wee need already labelled data, luckily, the folks at "cuentas claras" kind of described the process they took to flag the organisms in these documents. Taking this into account, there are two possible paths to continue working on this. We either "manually" label the data by processing it with a similar function to the one used by cuentas claras, or we try to apply clustering algorithms to identify patterns and groupings, and then see if we can notice any odd patterns that may indicate suspicious activity.
Or, we may try both approaches.

Manually labeling the data

Using this guide provided by the Open Contracting Partnership organization it should be possible (in theory) to determine if a certain purchase was 'suspicious,' as it provides a list with possible 'red flags,' indicating the conditions that a purchase must meet to raise suspicions.
After a brief investigation, we reached the conclusion that determining that case by case is not a good idea for two reasons:

• We are not qualified to analyze purchases and determine suspicious activities, since despite having that guide, we could never know with precision if it was truly a fraudulent purchase due to lack of context
• Labeling all that data is a very arduous and slow task.

Applying clustering

Although it wasn't the initial idea, by using clustering algorithms, we can study the information in different ways and eventually extract interesting conclusions. We might even be able to highlight certain purchases with interesting patterns... outliers.

Standard OCDS

The OCDS standard aims to store government procurement information in a clear, concise, and detailed manner. To achieve this, it provides a series of guides available on its website:

• Introduction to the standard
• Useful guide for understanding the data schema

After further investigation of the standard, we concluded that a procurement process is broadly composed of:

• A planning stage (Planning)
  • Information about the planning phase of the procurement process. This includes information related to the process of deciding what to contract, when, and how.
• A bidding stage (Tender)
  • The activities carried out to award a contract.
• An awarding stage (Awards)
  • Information about the awarding phase of the procurement process. There may be more than one award per procurement process; for example, because the contract is divided among different suppliers, or because it is an ongoing offer.
• A contracting stage(Contracts)
  • Information about the contract creation phase in the procurement process.

As you can see, this is not a simple process. It consists of several stages that we won't delve into deeply for practical reasons, but in an ideal scenario, we can conclude that the standard seeks to be thorough and comprehensive when managing this information.

Red flags

After studying and understanding the standard, we further investigated the famous 'red flags', where we found a useful guide to understanding the data schema which explains in detail several possible cases of fraudulent activities

From this guide, we highlight some examples.:

• The available time to apply for the tender is very short (less than 2 days)
• The tender does not have a clear criterion for selecting the applicant nor key information such as dates, budget, etc.
• The tender is by invitation only.

These are the clearest examples we can observe, but there are a total of 64 possible red flags 🤯

Taking these red flags into account along with the data available in the standard, we make a pre-selection of the data that interests us. This list can be seen in detail in the team's repository

¿What about N-1 relations?

Considering that RapidMiner works (initially) with CSV files, where all the information about a purchase is in one row, it would not be possible to process purchases that contain N-1 relationships, where, for example, there are several bidders with different data for each one. So to study this problem, it's best to start with the simple path, selecting examples that can be processed using RapidMiner. To deal with this, we have developed a script that iterates over all purchases in an OCDS-formatted file, summarizing the relevant information from purchases that are good candidates for processing.

This was not a simple task, as we first had to define what information about a purchase is relevant based on the red flags definitions by OCDS, and then deal with the schema to successfully extract it, resolving different conflicts along the way such as the use of different currencies to indicate prices, among other issues. Feel free to contact me for more details about it.

Uruguayan data

Once we've assembled a script to condense the data into a CSV with relevant information, we process the available data from the Uruguay site and observe the worst... while data has been stored in quantity since 2002, this data is incomplete, containing only information on the quantity of items purchased and their unit price.

Buenos Aires data

Although this seemed like the end of the project, we remembered that this is an international standard and many countries follow it. After conducting a brief investigation, we found on their website a map with the countries adhering to the standard, where we observe Argentina. We enter the Argentine data website, download the dataset, and after processing it, we observe that it contains many more data points than the Uruguayan dataset.

But... is it enough?

Duration of the bidding period vs. acquisition method

Conclusion

Future of the project

Interesting Websites with Information on the Topic: