 Welcome participants to the course on supply chain digitization. It is jointly being taught by Professor Piyanka Parma, Professor Sushmita Narayana and Professor Devapratadas from IIM Mumbai. So, in this lecture which is lecture 6 of module 3 and that is Analytics in Supply Chain Management. We are going to talk about specifically how AI ML can be used in demand forecasting and demand planning. So, let us summarize what we covered in last 5 lecture of module 3. So, we basically focused on what is analytics and then we observed what are the various characteristic of big data that is 6 bits volume, variety, velocity, variability value and veracity and we gave various reasons for this. Then we discussed what are the various types of analytics descriptive, diagnostic, predictive and prescriptive and then we gave numerous examples of analytics from supply chain management domain and we also specifically focused on one example called predictive maintenance. So, this was lectures 1 to 5 of module 3. So, in today's lecture we will focus on like why demand planning is important and towards the end we will discuss why AI ML is needed for better demand planning and forecasting. So, let us assume that this is your forecasted demand. So, we have applied some technique and tools and got this forecast. Then suppose when you realize the actual demand it comes out to be higher than the predicted demand. So, that means, your actual demand is more than the forecasted demand. So, what will happen? You will end up getting a loss sales. You will end up getting a loss sales. That means, some of the demand will not be able to meet it. Therefore, customers will go to their competitors and your demand will be lost. On the other hand, let us say your actual demand turns out to be lower than the predicted demand. So, what will happen? Then you will end up with inventory. So, you have to pay for the inventory holding cost and if the items are perishable, then not only you have to pay for the inventory holding cost, it might get spoiled also. So, therefore, if your actual demand is different from the forecasted demand, then either you will end up with loss sales or you will end up with having inventory. So, both the situations are not good. So, therefore, as a good supply chain planner, you have to make sure that your forecasted demand is close to the actual demand. And if you can do that, then your job as a supply chain planner, as a demand planner will be very good and satisfied, but that is not an easy task to do. Why it is not an easy task to do? So, for that we will recall the slide on big data. So, specifically big data has 6 characteristics, but we will take this 4 characteristics and we will try to explain why in today's scenario, like demand forecasting is a very important task and why it is very difficult task as well. If you look this various characteristic of big data, so let us first start with volume. So, volume means huge amount of data is being generated. So, for example, in today's scenario, a customer can place an order from any equation at any point of time during the day. They can place order from their home, they can place order after visiting a retail store. So, if you see the amount of data which is being created is huge. So, therefore, the volume of the data is huge. Now if you see the varieties of data, for example, customer review. So, in today's scenario, like customer can write a review that is in the text form about the product, they can describe in detail like what they liked about the product, if they do not like something about the product they can write in detail. So, many a time customers can take picture of let us say damaged product and they can upload it on company's portal. So, that will be an image form. So, the review can be in text form, review can be in image form. Sometimes customer review can be a number between 1 to 5, 1 represent very bad, 5 represents excellent. So, therefore, the variety of data can come from customer as a review which I need to incorporate in the model to get the forecasted demand. So, similarly if you look into the velocity data generated at high speed real time. Now unlike like 10, 15 years back when customer had to go to the retail store and buy the products, now customer can place an order online. So, they have the option of e-commerce, they have the option of Q-commerce, Omni-channel. So, various avenues are there to place an order and customers can place order any time during the day 24 by 7 and that data has to be captured. So, therefore, on a real time basis I need to capture the data from the customers and also there is another variability. If you look in today's time the customer behaviour is evolving. So, whatever they were liking may be 1 year back or 6 month back, the customer may not like the same product now because customers behaviour is changing. How customers behaviour is changing? That is being impacted by the social media presence that is being impacted by the digital marketing of the companies. So, customers are very much aware what is happening and they are very much knowledgeable now. So, they know which company is advertising for which product, what is the kind of promotional offer the competitors are giving, what is the price of the competitors. So, therefore, like customers demand pattern is changing is variable. So, I need to take into account that also. So, as a good demand planner, if I am able to capture the big amount of data, if I am able to capture the varieties in the data like some image form, text form, sometimes it is video form, audio form, in the velocity of the data variability of the data, then I would be able to get a good demand forecast. So, now if you look into this very carefully, so like volume of the data is huge, then there is varieties of the data like it is no longer a number, it can be image, it can be text, it can be audio, it can be video file and it is generating at a very faster rate and the data is variable also. So, therefore, like simple normal forecasting tools which we have been using may not work well because we have to capture all of these in the forecasting tool. So, therefore, AI and ML models play a good role and if I incorporate those models, I may be able to capture the variability in the data, I may be able to capture varieties of the data, I may be able to capture the volume as well as velocity and get a good demand forecast which could be almost same as the actual data. Now, let us spend some time and find out like what are the different AI ML models which are predominantly used for demand forecasting. So, we will list out few, so these are not the exhaustive list, but these are predominantly used by the demand planner across the companies, however, there can be few more also. So, let us spend some time to understand, so the first one is linear regression in which dependent variable is continuous, that means demand could be numbers. In this case, one assumption is the relationship between dependent variable and independent variable is linear. So, one example could be if I want to predict the number of units sold based on historical sales data, pricing information, then the promotional activities, demographic of customers, etc. So, it is a very common technique which is being used to predict the demand of customers based on linear regression model, but one of the assumption is that the relationship has to be linear. Then we have another model logistic regression, in this case dependent variable is categorical. So, I will give some example, then it will be there for you to understand. Let us take the example, I want to predict the probability of successful adoption of a new product. So, you are launching a new product, you want to find out whether the new product will be successful in the market or not. If it is successful, obviously there will be demand and I will get the revenue out of it. If it is not successful, I will not get the desired revenue. So, therefore, if I can develop a model and which will help me to predict whether the new product will be successfully adapted by the customers or not. Another example could be predict the probability of a customer making a purchase in response to a specific promotion or discount, because as a manager you have to make sure that you are at the top edge. So, therefore, many a time you may have to give promotion, you may have to give discount to bring the customers. So, if you are giving some discount, if you are giving some promotion offer, whether the customers are responding to that promotion or responding to the discount, how do I find out? So, for that I can develop a logistic regression model and I can predict what should be the probability that a customer will respond to my promotional activity and what is the probability they will not respond. So, after developing the algorithm, I will get to know what is the probability that they will respond to our promotional offer. If the probability is very high that they are responding to our promotional offer, therefore, my offer will make significant impact and obviously, it will have impact on our revenue. Then there is another model called decision tree. So, in this case dependent variable could be categorical or continuous. If it is categorical, then the decision tree is called classification tree. If the dependent variable is continuous, then it is called regression tree. So, decision tree can be used for both categorical dependent variable as well as continuous dependent variable and the name will change according to the type of the dependent variable. So, it can identify the key drivers or factors that contribute most to variation in demand. So, after applying decision tree, I will get to know like which factor or which feature is most important as far as my demand prediction is concerned. The good thing is that results of the decision tree are very easy to interpret and it is simple also. So, the manager or who cannot understand complex mathematics or AML model also, they would be also able to interpret the results and simple business rules can be derived based on the decision tree. The further it is very much effective when there are interaction and non-linear relationship among feature. So, if you see the various fees of big data, there would be non-linear relationship among feature, there will be interaction among features. So, therefore, if I want to incorporate this decision tree could be on such model. Now, another model is random forest. So, random forest is nothing, but extension of decision tree. So, I will explain. So, while decision tree is easy to interpret and understand, it may be sensitive to small changes in data and may sometime overfit to noise. So, decision tree has some issue regarding overfitting. So, therefore, n sample methods like random forest can address some of this limitation by combining the prediction of multiple decision tree. So, random forest is nothing, but an extension of decision tree where multiple decision trees results are combined and then predictions are done. So, random forest may remove the like possibility of overfitting. So, therefore, some many a time people prefer random forest over decision tree. Then we have another algorithm set of algorithm called boosting algorithm and the first one is Adaboost which is called adaptive boosting. The second one is called gradient boosting machines GBM, then we have XGBoost extreme gradient boosting, then we have light GBM, light gradient boosting machines and there are few more boosting algorithms are also there. So, these algorithms can handle large data set like including variety of features very efficiently. So, this can be adapted for dynamic forecasting ok. So, what happens nowadays like conditions are changing if you see the example of promotion. Suppose your competitors are giving promotion you have to keep an close eye to it. If your competitor gives promotion obviously customer will move towards the competitors and your sales will come down. So, your forecasted sale and actual sales will be different, but if I can capture this change in promotion plan by our competitors then my model will be able to incorporate that. So, I have to keep a close eye about what my competitors are doing. So, the one thing is regarding promotion. Another thing could be what my competitors pricing are. So, if they reduce the price obviously customer will go towards them. So, therefore, I have to make sure that I am constantly following what my customer strategy what my competitor strategy is and if they bring down the price I have to incorporate that into the forecasting model. The next thing could be economic change. So, in addition to the competitors strategy I have to also make sure that I am looking into the overall economy like how the overall economy is performing whether it is growing or not. If the economy grows and obviously customers will have money in their pocket and they will be able to spend it on some items and my demand will go up. If the economic is going down overall economy GDP is going down then customer will spend only on necessary item they will not end some items which are not necessarily or not requirement at this point. So, the demand will come down. So, all of these economic factors promotional strategy discount pricing of competitors have to be monitored continuously and I have to incorporate it into the model. In addition to this there can be some policy change also. So, government many a time brings policy changes. So, if I am not following what policy changes has happened from government point of view then definitely I will not incorporate it and it will have impact on my overall demand. So, I have to see how frequently customers behavior are changing how frequently competitors behavior are changing, how frequently competitors strategy are changing. So, accordingly I have to adjust myself and the model has to incorporate those adjustments. Therefore, I will get a good demand forecast which may be similar to the actual demand pattern. Then there are models related to the neural network. So, names are like recurrent neural network long short term memory LSTM in short then we have CNN convolutional neural network. So, all of these models can capture complex patterns and relationship in the data. So, as you have seen in the big data slide there are varieties of data and some are structured data some are unstructured data. So, for structured data it is easy to process and I can use it for forecasting purpose, but there are some data which are image data. As we discussed like customers review can be an in terms of image. So, they can send the photo of image product to you. Then it could be a video data also. So, customer may give a video feedback this upload it in the portal and then from there I have to see like what customers want what does not want and so on. Then feedback like then the data could be in audio format also and data could be in text format also. So, if I have this kind of unstructured data then neural network is a very good model which I should incorporate to get the demand forecast. So, as you have seen like depending upon the data type depending upon what are your goals of the forecasting tool we should select the model. So, these are few models which we have listed over here which are very frequently used IML models, but however we have to see how the demand pattern is changing how customers behavior is changing and what kind of data I have with me. So, looking into every aspect we have to finalize the model. So, before finalizing the model we can see like how others model are performing in terms of error in terms of accuracy. So, by looking into all of this phenomena we have to finalize a model. So, these are all IML based models which are frequently used in addition to this like typical forecasting models are there like moving average then exponential smoothing that also you can refer. But since our scope is only IML based forecasting model we are listing out this and in next subsequent lecture we will try to give some case study and then we will apply one of this model and show you like how IML model can be beneficial for demand forecasting. We will also see the error of this forecasting model and make sure that the model which are selected are having less error and we will try to incorporate that. So, there are various types of error which are used for measuring the performance of demand forecasting we will discuss those in the subsequent lecture. So, thank you for patiently hearing this and I am looking forward to your comments and feedback. So, kindly let us know like if you have any questions we will see you in the next class. Thank you.