Solve the Pricing(Quoting) Problem in a manufacturing company with machine learning algorithm
Data Science has so many use cases which spread across various industries and functionalities. Data Science has been used extensively in Retail, E-commerce, Health & Medical, Airlines, Telecommunications, Travel, Hotels, Sports, etc. Machine learning algorithms are solving problems for finance, HR, operation, supply chain , etc departments from long time. Today I am going to talk about two use cases i worked on while working in a manufacturing company in China.
So this company was in metal manufacturing industry. They used to manufacture prototypes for various companies from all over the world. This is not like typical mass manufacturing business where companies manufacture 100000 pieces of similar type. It was more sophisticated kind of manufacturing. They used to manufacture only 1–100 pieces and each part was handled with utmost care. Metals used were Titanium, magnesium, copper, zinc, Stainless steel or anything else. Company had a sales team from USA to Europe and Australia to south Africa. There reach was truly global.
Lets talk about the issues- They were facing two kind of problem. Problem number 1- was that the quoting prices was kind of a manual task there. They used to receive 3D CAD drawings from the prospecting customers. They had a dedicated team of mechanical engineers who used to access the drawings and then they will come up with the prices. But it was more of an educated guess. Dimensions, metal, quantity, etc were the most important factor in pricing. But these prices weren't uniform. If you give the same part to the 8 different people in the team. Everyone will come with the different price. Prices were either high or low. My challenge was — How to bring the uniformity there, how to make sure we don’t lose money in a part, and how to automate the quoting process ?
Second challenge was — Sales team used to receive so many enquires/drawings because of heavy marketing on paid and social both but the conversion rate for the company was around 27–28% on an average. I mean out of total 100 enquires which we quoted, only 27–28 of them used to place the final order and remaining 72–77 were just waste. Few customers used to take months in getting back. Just to give an idea about what kind of price range we are discussing here — Parts can be anything from 4000 USD to 30000 USD. Some parts were huge(OMOM) and few parts can be very small. So there used to be an open pipeline which no one had any idea about when it will get closed or how much they can expect in terms of revenue from the open enquiries. Company need to book the metals in advance, buy machines, hire people, expansion plans to new facilities, etc. Lot of things were depending on those open enquiries. And 27% conversion rate is not feasible. There was a dedicated team for quoting and sometimes the work load was way too much that there were delays in just quoting the prices.
We solved both of these problems with the help of two very simple algorithms. The most important thing in both of these cases was the data collection. I had no background in manufacturing and i had no idea where to start. So, I went very basic and started with dimensions(L*B*W, Volume and surface area) of the part and tried to quote the prices by using regression models. It was good starts but still quite far from the reality.
After lot of brain storming with our CEO, quoting team & manufacturing team we decided to use few other variables. We used quantity, metal , new/Old customer, Extra work on part other than manufacturing, how tough is the metal, density of the metal, etc. We managed to finally built a model which beaten the quoting team by miles. I basically followed a similar approach like house price prediction problem on kaggle. We used around 3 years of data. We built the model in python and deployed it for the end user. Its been more than 3 years now and teams are quite happy with the model. Its been optimised quite a few times now and we reduced the RMSE quite significantly.
For the second problem we used random forest algorithm and we predict the probability of winning an order. For this model we collected a huge amount of data just to understand what kind of customers are most likely to place order and in what time span? This model had an accuracy of 96% on training set and around 93.5% on test set which was acceptable. F1 score, precision and recall everything was above 89%. Sales team uses this model and they make decision based on the probability assigned to each enquiry. I even wrote a white paper for this solution and filed for a patent in Guangdong province in China and got a patent there. Below image is what we had after optimising it couple of times. It was okay but no perfect. Final Product had one magical ingredient which took the model to next level.
I had never imagined that problem of a manufacturing company can be solved by using the simplest algorithm and we can achieve such greta results. With the help of quoting algorithm the conversion rate of the company gone from 27% to 39%. There was a huge jump in sales and number of new customers. The prices quoted by the engineering team used to be high or low than the”true price” . So, either the customer of the company had to suffer. But with the help of quoting algorithm there was a uniformity in the prices. It was more systematic and no matter who uses the algorithm. Prices were always close to the “true price” of that part.
Thank you
M
