Price prediction in Healthcare Tendering

Predicting the final price of an auction has become an increasingly important area of business improvement for healthcare providers. Manufacturers can obtain a large profit by using previously predicted prices. 

At Konplik Health, we have over 6 years of experience delivering successful tendering projects for leading healthcare companies in more than 40 countries. From this, we can certainly voice that, at tendering opportunities in which price is a key criterion, both manufacturers and service providers in healthcare can increase their margins and revenues through predictive analytics. 

Predictive analytics uses statistics, advanced mathematics, and artificial intelligence to study large amounts of current data and past events so, patterns can be identified that allow forecasts to be generated for future events. It uses complex analytical models to analyze patterns in raw data and in turn, predict future results

Originally published at https://konplik.health/price-prediction-in-healthcare-tendering/


Automation and AI working together for tendering

Healthcare tendering is a challenging task for manufacturers of medicines, medical supplies, and equipment. 

The diagram below describes the steps we follow to deliver price-predictions at tendering. 

Benefits of process automation: data gathering, extraction, transformation and loading


On one hand, information gathering and managing processes can take weeks or months to complete. Thanks to advances in automation, however, a data-centric approach can often replace manual or semi-manual processes. 

What machines can automatically do:

  1. Data gathering

-Gather structured and unstructured information via scraping. 

  1. ETL (extraction, transformation and loading)

-Convert raw sources into a structured or semi-structured format.

-Combine public and private data

-Transform data with custom filtering, fuzzy product matching on large sets of data.

Modeling in price prediction

On the other hand, the lowest price is the key criterion for selecting the winning supplier at many healthcare tenders. Tender-based purchases are implemented to minimize the price for the duration of a contract. This practice is expected to reduce the cost as a result of price competition. 

Margins, as a result, remain under pressure due to the competing agents that bring about a never-ending price erosion. 

At tender opportunities that favor the lowest price, providers are required to submit a quotation. That is a fixed, non-negotiable price offer that cannot be changed once accepted. 

By collecting and analyzing past performance data from tenders, as well as current market prices and risk assessments, providers can take on a data-driven approach to bid more efficiently on market opportunities.

Creating predictive models requires more steps. As the diagram above shows, the train-test split is a technique for evaluating the performance of a machine learning algorithm. The procedure involves dividing a set of data into two subsets. The first subset is used to fit the model and is called the training data set. Contrarily, the second subset is not used to train the model.

One model doesn’t fit all

Not all prediction models relevant to prices in health tenders can be transferred from one market to another.

Despite the success of certain models in some markets, it is still incredibly important to be skeptical and question assumptions about the data and the analysis methods used.

Each tendering space is determined by different conditions that alter human decision making. One predictive analytics solution will not fit all. Empirical data proves this time and time again, as it frequently refutes carefully woven theories and models that work in other markets. 

In an upcoming post, we insist on a critical approach towards the blind application of game theory in predictive models of health tenders. 

Each market has its own rules for contracting, varies in the number and aggressiveness of competitors, and the availability of information from the players. As a result, our predictive models radically change from one market to another in order to obtain the best results.

We regularly test several models to find the best fit for each market behavior. 

Available information shapes competition

In our experience, the information available to competitors is the most dominant factor in shaping their decision-making.

In some markets, competitors have frequent and up-to-date information on bidding results. In Scandinavian markets, for example, healthcare tenders take place once or twice a month and competitors have precise information on how each player bid in previous opportunities. These conditions considerably limit the range of variability in bids placed by players.

In other markets, little or obscure information is the general rule.

In England or Wales, for example, the NHS intentionally withholds relevant information from competitors, so that their bids accelerate price erosion.

When a competitor loses in a bid, they are only informed of their approximate distance from the winning bid, and not the exact price that the bid won at. Likewise, when a bidder wins, they are also only given their approximate distance from the second-place bidder, precisely to get an even more moderate price in the next opportunity.

Evidence-based algorithms

The good news is that, while prediction models cannot always be applied to other contexts, methods often can.

We have successfully deployed tailor-made algorithms (including chunks of game-theory) along with machine learning algorithms (including gradient boosting, random forest, and Bayesian networks).

The key to obtaining a profit using previously predicted prices is to constantly check and review models with respect to the results obtained. 

On a regular basis, our algorithms automatically pitch a number of models against each other in simulations that decide which model was the fittest in the past, in order to subsequently decide which one to apply for the next bidding of a given product. 

For example, in some healthcare markets a buyer will purchase close to 2000 products at each tendering opportunity. The manufacturer we work for might bid for up to 50 products. For each of the products that are contracted through tender, we carry out a simulation to determine which algorithm would have obtained the best results in the past. 



Examples of algorithms

In the simulated competitions between algorithms, we have pitched algorithms such as: 

  • Tailor-made game theory-based algorithms. Stay tuned for a new post on this subject soon. 
  • Gradient boosting algorithms. Also, tailor-made. 
  • Tailor-made Time Series Models with some customization
  • Ridge Regression
  • Classification/Decision Trees
  • Random Forest
  • Support Vector Machine (SVM) for regression analysis
  • Bayesian networks
  • Neural Networks


The model that has obtained the best results for a given product, (this could be gradient boosting, for example), is the one we apply the next time to recommend a price for our client. A new simulated competition is carried out for each new product. This time a different model (Random Forest, for example) might prove to be the fittest in the simulation, and, as result, be the one that we deploy for the second product. 


Metrics used at price prediction

We regularly assess the performance of our algorithms using a variety of metrics that include

  • Margin: revenue minus the cost of goods
  • Revenue growth
  • Money left on the table: The difference between the submitted bid and the lowest-priced competing bids
  • Win ratio


Despite the success of certain models in some markets, we can say that each tendering space is different, and accordingly, one model will not fit all. Empirical data analysis is required to be able to increase the margin contribution and revenue growth through predictive analytics.