Webinar: Making every trade promotion efficient in the era of AI

so thank you everyone for joining us uh so today we're going to talk about a very interesting topic very challenging one um the promotion like managing promotion campaigns the efficiency of the promotion forecast how to be able to deal with that and so on so the topic of our webinar is making every trade promotion efficient in the era of AI uh so uh before we go to the like topic itself uh let I just wanted to introduce uh ourselves today the host of the webinars are Helen col she is the product director of the Leo AI inventory optimization solution hi Helen hi thank you very much for having me here and me uh my name is Anna I'm head of Business Development and Leo and I'm responsible for for communication with the customers with the existing customers and potential customers so thank you everyone for joining us me and

Helen will be together with you today uh so before we go to the like covering and talking about the promotion efficiency I just wanted to provide a couple of words to share a couple of words about who we are what we do and what our company is about so as Le we are specializing in working with the retail business uh for many many years we have a very good expertise in that and uh what are we do in a nutshell is we help retailers to uh organize automate and uh um manage their their different processes in the supply chain areas uh so regarding a couple of words regarding platform in general uh so Leo is a platform that combines like several modules the first one and um actually the the first like two we are going to focus on today is first of all it's like the inventory optimization block and it is the part that allow us to automate

the replenishment process at all the levels of the supply chain uh and uh be able to make like very precise order to allow the retailers to achieve from one side like their maximum service level maximum availability and from the other side not to spend like more money on investing into the stock uh the other part is uh it's shelf efficiency blog this is the solution that is intended to cover the entent merchandising process so to create their microspacing to create the store specific planogram to communicate the planogram to the level of the store for the execution with the help of the mobile application and return it back to the central office and of course I'm going analyze the efficiency of like what is placed on the Shelf uh together with the inventory and that's actually like what we are focusing uh today on is a promotion intelligence Block in our case of the leafl platform it runs like as a part of

the inventory optimization because like without the proper Automation in the inventory optimizations would be like almost impossible to automate their promotion processes so this is the blog that makes the promotional forecasting and helps to manage promotion campaigns in general from the side of the report management uh and last but not the least it's our brand new solution assortment performance so this is the one that is intended to um to help category managers to create the assortment matrixes to analyze their category to be able to make automatic clusterization of the stores by the formats and uh to analyze like each and every product inside their category to make sure like it's efficient and it's bring it's it is bringing out like Revenue oper profit uh okay so today uh like coming back to the agenda of like today's webinar the uh what we're going to to discuss today is first of all like the

main tasks for the inventory management uh like part and solution during their running like the promotion activities uh secondly it's promotional forecasting um how to build the efficient process using this promotional forecasting how to choose the right model that will work for us uh how to uh like build and check some hypotheses and uh to like how to choose and add forecasting factors to the model uh afterwards of course like after like the promo campaign finished it's very important to make the evaluation of the promo campaign to be able to make some um like uh some uh summar some conclusions for the future promotion campaign so we're going to talk about the evaluation and of course and like we're going to share like we've been in that like for several years uh recently and we have a lot of experience we have done a lot in this sphere so we we're

going to share some like lesson to learn lessons to learn um regarding managing and uh building promotion campaigns so what's important to mention here is that today we're going to share like our our experience in doing that uh and uh of course it's like today we're going to talk about some general things if you're interested in this topic in particular we will be glad to arrange additional like demos for you and all like meetings to talk about that like more specifics uh so uh let's uh before we go like to the discussion of like how to make it efficient I would like just to highlight what is like what's the problem in promo like what's the dilemma with the promo why it is so hard for everyone and about from the other side why is so like important for everyone so what we seeing like working with the uh retailers recently that the uh percentage of the promotion campaign increases Year from year and not only

like in the grocery retailer business but also like in different spheres so in order to like increase the Loyalty in order like to keep the customer like come into our stores like we need to like be be in Trend and we need to suggest some otion activities so uh that's why promotion is like a necessary AAL for every retail business so to be able to be like on the level with the other retailers because like everyone is doing that and that means that like we cannot be the one who doesn't do that uh and uh everyone like who faced somehow like with the promotion knows that it's very like exhausting process it takes a lot of resources and moreover it um during like planning the promotion campaign a lot of Departments of the uh a lot of departments from uh the company is um like involved in that so how the promotion uh how the like how the promotion is structured how it is build

so first of all like uh uh the process looks as following first of all we need to plan the promotion campaigns after that to prepare for it after that of course like to ensure the execution uh and uh to make sure that we are doing their like uh analysis of the promotion campaign and to exit promotion campaign so the first stage is like very critical and the most like exhausting one so because it includes like first of all the to select the supplier sometimes it is initiated by the supplier to choose the promotion mechanics uh what's like the most like difficult here the most difficult part is to make the promotional forecasting so to understand like what increase in sales we're going to have uh and it is like uh very hard to do when like there are like hundred or thousand skus that are like running in the promotion campaigns afterwards negotiate these conditions with the suppliers approve with them and after that we can start the promotion campaign itself uh so uh can we go yeah thank you

very much so what's like the difficulty like before that we are talking about like some general things but if you talk about the particular the area we are we are expert and we are focusing on uh like the replenishment side uh like what are the difficulties here is first of all the promotional forecast it it requires like high accuracy the sales can increase like three 10 50 times and uh that's why like it's very important to make the like rather precise and good forcasting because like out of stocks they it's like very critical for the promotion campaigns especially in the beginning and uh so like there are two main Pains of managing promotion campaign from the side of replenishment so first of all it's making the forast to make sure that we have the highest availability level in the beginning of the promotion campaign and during promotion campaign and the other paying is that we are not going to the overstocks in the end of

the promotion campaign uh okay so what is um yeah as I mentioned it's like the main task for us and from what we see it's like the most like one of the most complicated things to do because uh like no one knows like what what the increase is going to be like in sales during the upcoming promotion campaign and sometimes uh we need like to prepare before the promotion campaign for example for some huge promotion periods like uh the Christmas time there are like several promotion campaigns that are running we need to prepare our like human resources for that we need to prepare our stores for that to make additional displays and so on but of course the main thing is to like ensure they're like very good and very precise stock level uh during the pro promo uh so uh what is uh uh like if we talking about the uh promo like what are the actions that need to be taken uh for

like running the promotion campaign so first first of all we need to uh uh calculate and to forecast the accurate like up in sales to be able to make this like initial order and keep the uh demand during the promotion campaign and to plan the promotional like display size for example like there can be some like I don't know some additional display of uh some champagne in the store and so on secondly uh we need to identify and uh know like when to make the purchase order uh because like there are like different lead times there are different suppliers with different schedules and some of the uh like some of the goods they coming like directly to the from the suppliers to the store and some of them they are received for the in like to the central workhouse and afterwards they need to be delivered to the stores as well so it makes this like even more complicated and we need to make it like some time before the promo uh uh the other thing is uh how to uh um

how to like before the promotion starts we need to give some time to our uh store managers like store personal store employees to make the merchandise and layouts to change the price taxs to make some promotional like science and so on and after like in parallel we need to make sure that we uh are fulfilling the store so like when the promo starts and like before the promo starts they should have already like their uh uh these Goods like available in the stores uh and uh of course like we uh when the promotion finishes it's like a typical pain for the retailers that there are a lot of like overstocks there a lot of like money Frozen in the overstocks and of course when promotion finishes the price goes down goes like to the regular one goes up uh like no one is going to like buy it there's like no advertising and so on so it will be hard to get rid of these like overstocks and um summarizing like all of that of course

it's very like it's crucial to make the ongoing analysis of each and promotion campaign after it's finished and to like make some lesson learns for improving this in future so uh uh I'm giving the word right now to Helen she's going to talk about like the the most interesting part like the forecasting itself how to make it our experience and we hope that this part will be like useful for you and you can take something for like your planning promotion activities thank you very much and so yes now we are going to talk about the forecasting and some useful hints that uh can be relevant in case you are uh planning or you are starting or you are just in the middle of this process uh from our experience so um regarding the flow of working with uh uh such kind of promotional activities forecasting the sales up

leaves forecasting uh we can divide it into three steps the first step is the data preparation so uh I think that everyone knows that the trash data in trash uh forecast out and um it is crucially important to make the data preparation uh at the very beginning and the initial stage of the making this activities with the uh forecasting uh and especially forecasting of the promotional sales uh so on this step um from our experience it is important to get rid of the out of stocks and out of shelf uh days so yeah we are dividing the out of stocks where we have zero balance and we identify automatically the out of shelf days when we assume that uh we have the positive quantity on hand but the sales are not so relevant that's why probably we have some issues with the layout on the sh on the shelves

uh we also get rid from the spikes in demand and it is important to clear and divide uh the historical sales into their Promotional and regular sales to have like the basement for these calculations on the appropriate data the second step is identifying the factors uh that can influence the promotional demand uh especially comparing with the non-promotional regular demand and for each and every uh retail sector indust uh these factors can be different and even in terms of the same industry depending on the region for example this factors can differ as well and afterwards when we are identifying the list of these factors that are like the hypothesis for us we are starting the process of checking it is called the cross validation uh process when we are training the model on the data set afterwards we are making the validation

so basically we are training on the data afterwards we're feeding the model with that data that is not known to the model and afterwards uh we are like trying to understand the efficiency and probably make some changes in the model uh in the in the number of factors in the weight of these factors and so on and afterwards we are making the train validation test that helps us to understand the general forecast and accuracy and deficiency of the forecast uh so if talking about the different types of models that can be used for this specific task uh there can be like the very huge variety of these models uh it is like the general examples of models that can be used uh for the promotional forecasting uh so among them you can see time serious forecasting models with linear dependencies with recursive dependencies and and even the

decision tree models that are most commonly used for precise and efficient promotional forecasting um actions uh so um and in some cases there can be a couple of models that are used while identifying the best one for each specific case several year years ago uh we were analyzing the results of so-called M competitions uh M competition are a series of open competitions to evaluate and compare the accuracy of different forecasting methods and uh data scientists all over the world uh they are trying to compete in this competition and to win some prizes uh they have started their competition in 1982 and uh basically the competition is focused on uh different forecasting tasks uh depending on the data set uh

that is relevant for this specific competition so in 2012 there was a so-called M5 competition where the task was to forecast sales for Walmart who generously made available their data set and the uh first prize uh went to the data scientist team uh who was using um a couple of different models but the general the most uh common L used model was light GBM and just now lately due to our experience as well we are using light GBM model uh for like the biggest variety of cases while we are making the promotional forecasting for our customers uh regarding the some useful hints in this process um when we are trying to evaluate the best possible model uh which need to be used uh in some specific cases we need to uh

clearly understand the aim of the forecast uh because dependent on the aim different models can be more or less accurate U because of their specifics so uh of course we need to understand whether we are trying to forecast the Fulfillment of the stores during the promotional campaign or the Fulfillment of the distributional center or we are like trying to make some Financial Planning and depending on this uh different models can be more accurate than other ones uh so if the uh we are forecasting different categories with different factors that are influencing the forecast um we can see that uh there are some dependencies with the weight of different factors in terms of different categories and uh here is like an example how this weight can be identified and while we are identifying it for our customers while we are doing forecasting for them also uh it is

important to understand the general aim and u based on this uh we need to identify whether we are making the top down or down up forecasting approach so whether we are trying to make the precise forecast on the SKU location level and afterwards we are going up to the level of the SKU and up to the level of the category for example uh and usually we are making such approach when we are forecasting the Fulfillment of the stores and upside down um when we are making the forecast on the level of the category and afterwards we are um making the down up approach while we are going to the SKU level and afterwards to the SKU location level to uh and starting from this approach probably the main aim here it can be a little bit different uh when when we are trying to make some forecasting to share them with our vendors and make some financial

planning as well uh also we need to understand if the promotion uh is going to be done throughout the whole retail chain or for the limited number of stores and formats we had an interesting example lately with our customer who had five different formats of their stores and some particular es was introduced into only one format of the stores uh which uh included if I'm not mistaken 15 storage locations so when uh uh it was introduced only to these 15 Stores um and the forecasting the promotional forecasting was done only in terms of these stores and they had some historical data uh our customer decided to introduce this SKU into the other storage locations as well uh but there were no historical dat for the other stores for this specific SKU and there are some specific models that can take it into account and uh uh make the

forecasting uh taking in mind that the S was introduced for limited uh time um of past only for some specific stores uh also it is important to understand the fullness of the data set and uh understand that for example some more complicated models they need more wide range of data and some more simple uh models they can work well um good enough let's say with the limited quantity of the data so it depends on the Range the wideness of the range of the data set and of course the promotional mechanics so when we are talking about the percentage of the discount it is the one type of models can be used where the price factor is one of the most important one and when we are talking about 1+ one 2 plus one or other more complicated mechanics uh this factor is important but some other

factors can be more important and it will depend on the efficiency of the model so this must be taken in mind as well so if talking about the data that uh is needed to make the forecasting uh of course for once more time it depends on your aim but when we are talking about the forecasting of the uplift when we are talking about the uh forecasting of the sales during the promotional campaigns um usually we uh need this type of the data uh when uh we take the data of the historical data for the names of the promotional campaigns the promotional type promotional mechanics uh their duration of the promotional Campaign which is crucially important uh the promotional advertising of course the Le of the skus that were um or are going to be in the upcoming promotional campaign uh the price of course and the calculated percentage of the discount

the storage location uh some additional layout by the way the additional layout or additional displays can influence a lot the forecasted sales because uh if you're increasing the presence of this SKU in terms of the storage location it will definitely increase the sales uh of this SQ uh the category hierarchy uh and uh the sales during uh and before their promotional campaign started uh so um about the evaluation and measuring the efficiency of the promotional campaigns um it is like an important Point how to understand whether this promotional campaign was effective whether it um met the aim for which uh all these efforts were made during the uh preparation stage during uh the evaluation stage and everything else so we need to clearly understand if

everything was okay so from uh the point of view of the forecast in accuracy uh it can be done on different levels uh of their retail um retail business so it is a very I assume it is the most commonly uh asked question uh what is your forecasting accuracy what is the level of your forecasting accuracy and uh usually it is a very complicated question for us uh because it can be measured with different um indicators it can be measured on different levels so for example if you're are measuring the forting accuracy on the SQ location level of course it will be much uh much less than the forecasting accuracy on the Enterprise level for example because we're aggregating the data so depending on the aim of the forecast we need to identify on which level we will make the measurement of the forecasting accuracy

if we are talking about the task of fulfillment of uh stores during the promotional campaign and at the beginning of the promotional campaign uh usually we are measuring it on the SQ location level because it is crucially important for us to understand whether we were efficient enough to cover the uh uplift of the demand during this very interesting and unstable periods of time uh regarding the metrics uh there is a wide variety of the metrics that can be used to understand the uh efficiency and the accuracy of the um promotion of the forecasting overall but uh in um forecasting accuracy in terms of the promotional campaign as well uh so there are just an examples of such kind of metrics that can be used usually when we are measuring the accuracy of the forecast on the level of the SQ location for the upcoming promotional campaigns we're using w m

weighted map indicator because uh the price is taken into account in this perspective and under forecasting of sales for some very uh cheap SKU and at the same time under forecasting the SKU uh with the very high price with expensive uh expensive price it is like different uh different influence on the overall uh specifics of sales uh and uh a lot of companies for example in the M competition I was mentioning previously they have their specific and another metrics that was used to uh calculate the forecasting accuracy uh but what is more important for us uh it is to understand and to measure the financial metrics and outcomes uh after the promotional campaign ended even if you have 100% uh accurate forecast uh you can have some another issues that can be connected with the suppliers Logistics uh with the

on store operations and Etc that can decrease the overall efficiency of the promotional campaign that can lead to the loss sales during the promotional Campaign which is just very bad situation uh for any retailer I assume so we are more focused um with our customers we're are not more focused to understand these specific indicators uh we are understanding the uh actual sales difference between the uh between the uh forecasted sales uh we compare between each other the promot post promotional overstocks the post promotional inventory turnover and the percentage of the Ava liability and I'm giving the word to Anna to talk about one of our uh great customers who are using our approaches in terms of the forecasting and which

outcomes they're getting uh while using it they right now they have around like 90 location they have a huge percentage of their assortment in promotion um about 70% of the assortment it is replenished from the distribution center and uh they have four different promotion types and they manage in general 30 30,000 skus uh so uh uh let's uh uh like let's take a look how it it was going to H how it was happening like on their side uh so first of all uh as Helen mentioned like it's very in general it's very important to get the data and that was like the first step that we did with nov so we collected the data it's very like kind of exhausting process because like for to be able for the model to run we need like a lot of data we need like the maximum data noas had this data luckily

and but still we needed like to prepare that to like see what's missing there and so on so as a next part and um we made the modeling and um like we made some like different kind of models and after that we like needed to understand what and to choose what was like the most efficient model and afterwards like to implement this model in this particular case from the reality of this particular retailer uh so we did the modeling and we can uh like we compared like the models that like the results that we like received with the for C that was done by the customer and which choose like the most uh the most like relevant the most efficient model after that uh when like we already like know what model we're going to use we go to the pilot so we do that on the limited assortment uh like kind of like to approve our um to approve what we did uh and afterwards like when we are sure that okay

everything work smoothly like we need we made some additional like settings and so on we are doing the scaling for their like all promotion campaigns in the company uh so let's take a look like let's take a look what exactly like like what specific results were achieved by noos in this case uh so uh due to their uh like usage of our methodology our like promotional block the availability increased by 2% and I would like to mention here that even before I mean for the promo products even before that the availability level of this particular customer it was already on a high level and uh we even managed like to increase that like almost to the maximum one so and the other uh uh the other part that we managed to decrease overstocks two times like overstocks that are um that are happening after the promotion finishes so we decrease it like two

times and help the company to release some money for like future development and Future activities uh so uh in general uh what kind of business value what kind economic effect to expect from this kind of solution is first of all it's up to 20% of the improved availability depending on the company depending on the retail vertical and so on and up to 50% decrease the overstocks and release money for you for like some some additional activities some new stores opening and so on uh what are the other effects uh of like using the promotion automation tool it's like less time and resources that is needed for like planning the promotion campaign and um for planning and preparing the promotion campaign and preparing the replenishment itself during the promotion campaign secondly that uh all the trade promotion history is centralized it is stored in Leo and

we have like the whole all history for like all the promos that were run and this is critical for the uh machine learning model and like the more history we have the more like the model like learns on on it and improves the forecast in accuracy so are there are like very good results in general like to be expected from this kind of solution and uh automation of the promotion campaign forecasting uh regarding the lesson that were learned by US during uh our experience with working with a lot of customers in terms of the forecasting uh of the promotional campaigns in particular um we can divide them into two groups uh in forecasting area and in process area uh because uh generally we are making this precise forecastings calculations and everything to take all the task for the

most efficient um business processes in terms of the general uh overall promotional uh business process inside the retail company so we are focused on the forecasting of course it is crucially important but still we are taking into account and consider all uh other processes that are touching uh the promotional campaigns so in terms of the forecasting uh it is impossible to have the accurate forecast without clear understanding the main aim of the forecast so it was described uh in details when I was talking about the um making the decision on the particular model for making the forecast it is impossible to calculate the accurate forecast without full and well-prepared data and uh from our experience I can share that um we were trying to increase the forecasting accuracy with some adding some factors and some I don't

know some more complicated logic using some more complicated models and Etc but uh like two years ago we understood or three years ago we understood that um we can uh really influence the accuracy of the promotional campaign when we are preparing the data very accurately uh and it really uh improves and uh have the big influence on the accuracy of the forecast some models can have the same forecasting accuracy in the initial stage but in the Strategic perspective we face and see that machine learning models uh they show the better results because they are learning on their own experience that is anality is crucial factor for all the categories even if it doesn't seem so uh and we will talk about it um both with the price elasticity and I will give an example of this so and the price elasticity is one of the key factor that influenc in the

um forecast and the forecasting accuracy must be analyzed only considering some other factors and I have an interesting example from uh um one of our customers uh so we were analyzing how uh the price influence sales so the price elasticity and we saw that for the specific uh ASU it is bananas when when the price is going up the sales are going up as well it is a very strange situation and uh we need to understand like more why it was happening and we decided to understand the seasonality of bananas and we can see that in Europe it is the our customer from Europe uh so in Europe we see that uh during some uh periods of the Year especially the summer perod period and the late spring periods the sales are

decreasing because they're cannibalizing by the sales of another uh seasonal fruits and vegetables and when we are adding The Cure uh for the price we can easily understand that uh the price goes down when the demand is going down and the price goes up during the seasonal fluctuations so when we are talking about the forecasting know the price elasticity analysis we need to take into account the seasonality for example in for this particular case but of course there can be some another cases uh that can be relevant for some other product categories in terms of the processes uh so the accuracy of the forecasting is very important and we need to measure it precisely but the financial outcomes are even more important uh it is very important important to rely on actual promotional sales rather than on the forecasted

sales uh so when we are making the forecast in advance for the first order to fulfill uh the beginning of the promotional campaigns of course we will rely on the forecasted sales but when we already have the actual sales data we need to rely on this data as well the forecast is a part uh so we must be we must be integrated into the general promotional planning stage and uh we as I have mentioned so when we are making this forecasting uh we need to integrate it we need to um make it like an the same process for the planning of the promotional campaigns it is important to check the efficiency of each and every Promotional and compare the aims with the outcomes considering all the factors that were making uh the contribution uh in one of the most important so um generally uh we need to make this um like comparison between the aims and

the outcomes and understand what was wrong and precise finish of the campaign is as important at the precise beginning not to lead to the high level of the overstocks we need to uh like um generally understand how to work with some addition layouts for example and some other factor that can influence this Factor as well so I guess it is pretty all from our side in case of any questions it will be great to answer them and uh for once more time uh it is a very interesting topic for us for our team uh we have great experience in this and we can share some additional information uh with you if you have some specific questions on your specific processes thank you very much everyone for the attending our webinar if you have an interest in this topic we will

be glad to talk into some more details