Questions/Answers

Base module

How to understand why the system predicted specific quantity?

• To do this, go to the Analysis form and start calculating the forecast for interested position. The main and most effective forecast check is visual. It is recommended to visually evaluate the sales dynamics of current and previous year, comparing it with the current forecast in order to determine the use of seasonality in the forecast. You can also add a price to the chart to evaluate the effect of the price on the forecast and on sales. You can also use visual analysis of other factors.

Why did the forecast differ a day ago?

• Because the system recalculates forecasts everyday to take into account the latest sales data

How to understand what forecast was a week ago?

• In the Analysis form, in the parameter selection area, click the settings button, and then select the required last week of sales in the week field, which the system will see when making a forecast.

How does the system predict new positions?

• There are two methods: with an analogue, or based on sales of an average position in the group. If system has an analogue, then the sales history of one or several analogues, multiplied by a coefficient, is used. If forecast is based on sales of an average position in the group, then the sales of the group are divided by the number of positions in it, after which they are adjusted for the price elasticity of the group with the price of a specific new item.

If the position was sold in some stores and its distribution is expanding, how will the system predict it?

• The forecast for the expansion of distribution is based on sales at the group-store level, sales at the group-region level and sales at the SKU-region level. To simplify, the proportion is taken: [SKU-store forecast] = [SKU-region forecast] * [group-store sales] / [group-region sales]

How many weeks does it take to use analogue sales data for a new position? How many weeks does the system need so that it can build a forecast based on sales data for a new position?

• It is recommended to use analogue sales data from 6 to 13 weeks, depending on the system settings. However, this does not mean that for 6-13 weeks the system will not see data on the actual sales of the new position. As soon as the first data on the sales of new items appear, the system will already use it in the forecast, and the sales data for the analogue will only be used for those periods where there were no sales. The system can make a forecast for a new product without an analogue even when there are only a few weeks of sales, however, to more accurately determine the dependencies of sales on various factors (seasonality, price, promo), it is recommended to allow the system to use analogue sales data for a longer period of time

How does the system forecast new stores?

• The forecast for new stores is based only on one or more analogues (similar stores). As with new positions, it is recommended for new stores to allow the system to use sales data for similar stores for 9-13 weeks, so that the system can more accurately determine the dependence of sales on various factors. However, after the first weeks of sales appear, the system will see the sales data for the new store and use them in the forecast. The system will use data on sales of analogue stores only for those periods where there are no sales / stocks of a new store

What will happen if you do not add an analogue for a new store?

• For the first week of sales, the system will not build a forecast at all, but after the first week of sales, the system will build a forecast without taking into account the dependencies, based only on the sales of the first weeks. With the accumulation of history for the new store, the forecast will become more accurate. However, it is recommended to add analogues for new stores.

What is the formula for calculating the forecast value? How to check the forecast value?

The total forecast for the week is equal to the sum of all its components. Prediction components can be viewed in the Analysis form by selecting them in the "Components - In pieces" menu. Example: the forecast is built on the basis of the actual price, seasonality and the base level. We choose the predictors “A. Price”, “Seasonality”, “Base” from the menu “Components - In pieces”, summarize, their sum for each week will be equal to the forecast value. Example: the value "A. Price" = 20, "Seasonality" = 25, "Base" = 100. The forecast value will be equal to 145.

How to understand why the forecast is not equal to sales?

The situation when the forecast is equal to sales - its almost impossible. However, if the differences in forecast and sales are not significant and are covered by a small safety stock which is calculated by the system, then this situation is normal. If the differences are significant, then it is necessary to try to determine which factor is not taken into account in the forecast. The algorithm of actions is as follows:

• Check the forecast at the SKU-region level for region 0 (the entire chain), either at the level of a specific region or at the store level

• In the "Models" window, determine the factors that have the highest correlation coefficients

• Check if these factors are taken into account in the forecast

• If all the significant factors are taken into account, it is necessary to analyze and find out which factor the system does not know about

• If not all the significant factors are taken into account, then it is necessary to add these factors to the chart and evaluate visually whether their influence is fully taken into account

• If you can’t understand for yourself whether all significant factors have been taken into account or not, contact MySales for analytical support.

Why, if the differences in forecast and sales are not significant, then the situation is normal? We will lose sales if the forecast is lower.

The last statement is not true, since there will be sales losses only if the sales exceed the forecast and the safety stock in total. If sales exceed the forecast, but remain within the safety stock, there will be no loss of sales.

How is safety stock calculated? How many percent does it reach?

Safety stock is calculated individually for each item in each store. Safety stock is the standard deviation of past sales from the forecast. However, in practice, the system always tries to make it more accurate, counting on different forecast ranges. For this, the forecast is divided into 3 ranges:

• High range - promotional periods and periods of high seasonal sales

• Low range - periods of seasonal decline or periods without a promotion, if the position has a slightly expressed seasonality

• Normal range - other periods

Further, the standard deviation is considered separately for each of these periods. Thus, the safety stock, for example, for ice cream or kvass in winter and summer will be different. Safety stock for promotional periods and non-promotional periods will also be different.

The percentage of safety stock in relation to the forecast will also be different for different positions, but the general rule is as follows:

• The higher the speed of sales, the greater the mass of demand for it, the more accurate the forecast, respectively, the lower the percentage of safety stock

• The slower the position is sold, the greater its safety stock in percent, but at the same time less in pieces. To understand this, imagine a simple case where the forecast is equal to average sales. If on average, a position is sold 0.5 units per week (every 2 weeks), then its safety stock will be slightly less than 0.5 in units, but at the same time, it will be almost 100% of the forecast.

Can insurance stock not cover all 100% of sales peaks? What to do in this case?

It could be. However, this can only be a problem for the top, best-selling items. For such positions, we recommend increasing the safety stock using SS coefficients. In this case, the safety stock will be multiplied by this coefficients. For non-top positions, it is recommended to leave the SS coefficient by default. Usually it is equal to one, this is a system setting and it can be changed.

What if the safety stock is too large and could potentially lead to write-offs of perishable goods?

Since sales of specific positions are often volatile, and the lower the sales rate, the higher the demand volatility, as a rule, it is important to find the right balance between the availability of goods to cover the extremes and the level of write-offs. The general recommendation is to reduce the SS coefficient for such positions and set it to less than one. And for top positions, where the delivery time does not significantly exceed the sales period of position, set this coefficient close to zero.

Do these ratios need to be reviewed regularly?

This just does not need to be done. By adjusting the coefficients, you accumulate experience, and you should not constantly change them. For more information on the application of SS coefficients, see the section "Safety Stock adjustments".

In what cases do you need to enter a master forecast and how does it work?

If the Promo module is not used, it is recommended to enter a master forecast for promotional periods. You can also use the master forecast to adjust the forecast for new products, if you have not set an analog. The master forecast is introduced at a general level (SKU-all stores) and distributed proportionally to the system forecast at the level of specific stores, raising or lowering the total forecast of all stores to the level of the master forecast.

What is region 0?

Region 0 presents aggregated sales data for all stores in the chain. For region 0, a separate forecast is being built, which may differ from the sum of forecasts for all stores. The level "region 0" is used as the highest level of the geographic hierarchy when building forecasts of lower levels.

What is the difference between the “Sum stores - SKU” and “All stores - SKU” levels in the “Forecasts - View” form?

The level "Sum stores - SKU" is a forecast built at the level of each store and summed up. The “All stores - SKU” level is a forecast built at the level of region 0, which represents aggregated data for the entire chain.

Why are the forecasts built for each store and summed up different from the forecast at the region 0 level?

The difference is that when building a forecast at the store level, the system analyzes the sales dependencies on influencing factors at detailed levels, which in certain cases, when there is enough data, can give a more accurate result, and in some cases the dependencies may not be determined.

Which forecast is more accurate, at the regional level 0, or the total forecast for all stores?

For different positions - in different ways, it depends on is it enough sales history at detailed levels to make a forecast. With insufficient history, a forecast at region 0 level may be more accurate. Another advantage of using the forecast at the region 0 level is that it allows you to determine the chain-wide sales dependence on influencing factors (price, seasonality, etc.) and use them at more detailed levels.

Does the company need to organize systematic work with the functionality of "Anomalies and Excluded"?

This is not necessary. The Anomalies and Excluded functionality is used in exceptional cases if anomaly sales are identified in the forecast analysis process. Also, if the customer’s system has dedicated sales by reservation, it is recommended to download them so that the engine takes them into account as anomalies. In some cases, if the promo module is not used and there are requirements for the engine to predict regular and seasonal sales, those weeks for which the master forecast with the Promo type is set are automatically pulled into the forecast exceptions.

For which positions is it worth using forecasts built using a neural network?

The list of such items is configured in the system. Before you use ф neural network to forecast a position, you need to verify the forecast, i.e. check visually. This can be done using the “Brain” forecast profile.

In our stores, sales volumes are changing, for example, in connection with the opening or closing of a competing store. How quickly will MySales respond to such changes?

Most likely, the reaction of the system will not be instant. This is done in order to smooth out local extremes in sales. But after 1-3 weeks, the system will see these changes and begin to take them into account in the future forecast. This can be checked on real, already happened examples, rolling back in the forecast 1-3 weeks ago. You can do this in the "Analysis" form by clicking on the "Settings" link and choosing the last week of sales that the system can see when building the forecast. Also, immediately after the first week of increased sales, the system will begin to raise the safety stock, doing it carefully and without creating an increased stock where it is only one such week in history

What if, after analyzing the forecast, I do not agree with it?

In this case, there are several options:

• Check other forecast models that the system has built and select the most suitable. You can check the models in the "Analysis" form by selecting them from the "Components - Analysis" menu and checking visually on the chart. It is recommended to start from more accurate to less accurate. Further, the desired model can be fixed for the engine by clicking the "Models" button and selecting it there, however, it is recommended to do this at specific levels (SKU-region, SKU-store, group-region, group-store).

• Correct the forecast of the system using the "Master" functionality at the SKU-all stores level

• In exceptional cases, you can download a detailed forecast at the SKU-store level

• You can also request analytic support from MySales.

How can MySales automatic forecast be useful for ordering fashion group items (clothes, shoes)?

Clothing and shoes that are ordered on a regular schedule can be ordered based on the system’s forecast. If the purchase of clothes and shoes is done for each season, then the organization of this process should be built a little differently. Initially, it is worth using the forecast for groups of models (for example, men's sports shoes), in order to determine the potential demand for them. Next, you need to expertly select models for potential demand and also expertly make a decomposition of the forecast at the level of model groups to specific models, and then to the dimensional grid to get specific numbers of positions for the order.

Promo module

What is a promotional uplift?

Promotional uplift is an additional increase in sales during the promotional period, which occurs due to communications, when more buyers knew that a certain position has promo offer for a limited period of time.

Why is price elasticity insufficient to predict a promo?

There are two reasons. Firstly, if you just lower the price in and don't tell anyone about it, it will not give such an effect in sales growth as reducing the temporary price with the support of additional communication. Secondly, often a promotion is held without any replacement of price tags, when the discount is available at the cash desk, and the buyer learns about it from additional materials: a leaflet, an electronic catalog, a banner, etc.

How does the system predict the effect of communication in a promo?

The increase in sales due to the communication effect is the most difficult component in forecasting. Firstly, this effect is very different for different positions. Secondly, for specific positions, there is often lack of promo history, since in retail there is a significant rotation of the assortment. Prediction of promotional growth in the system can be divided into several main stages:

• Analysis and allocation of historical promotions for all SKUs

• Training system for matching comparable promotions

• Neural network training for predicting promotional uplift, if it was not possible to pick up comparable promos

• The selection of comparable promos to predict uplift for future promos at the SKU - all stores level

• Prediction of the promo uplift coefficient at the SKU - all stores level using a neural network, if comparable promos could not be found

• Adaptation of promotion uplift at the level of specific stores, based on the history of the promotion for each position at the SKU-store level

There is a feeling that we are giving a discount above the psychological level when the consumer is ready to buy. How to understand this?

To understand this you should simulate several promo scenarios, with a different discount level and various communications. In order to simulate a promo, you can either create a new promo, with the conditions that you want to simulate, or a new variant of an existing promo. You should consider that when creating a new variant in an existing promotion, a discount that is determined individually for the items in the promotion has priority over a discount that is determined at the level of the entire promotion, therefore, if different discounts for each product are defined in the promotion, it is better for modeling a new scenario to create a separate promo without approving it.

In the form where promo forecasts are displayed, there is a forecast called "Base" and "Promo". What do they mean and how do they differ?

Base - a basic forecast without taking into account the effect of communications, but taking into account price elasticity. It shows how much you would sell if you simply reduced the price and supported it with basic communication in the form of an indication on the price tag. Promo - a promotional forecast taking into account the promotional uplift due to all the additional communications in the promo, allowing more customers to find out about a profitable offer.

How to verify the promotional forecast for a SKU to make sure that the forecast is correct?

First, you need to check the comparable promos that the system picked up. To do this, press the "C" button next to the SKU to see comparable promo SKUs, sales on them, as well as their uplift rates. The forecast may differ from sales of comparable promos, as conditions change: price, discount, season, cannibalization. In order to analyze the factors affecting the forecast, in more detail you need to double-click on a SKU in a promo to go to the "Analysis" form and see in more detail all the components of the forecast.

In the form where comparable promos are displayed, there are fields: "Before,%" and "After,%". What do they mean and how do they differ?

"Before, %" is the coefficient of promotional uplift to the forecast, which does not take into account price elasticity. "After, %" is the coefficient of promotional uplift to the forecast, taking into account price elasticity. Typically, "After,%"" is lower than "Before,%". The system uses these basic coefficients from comparable promos, applying them for forecasted promos at the SKU - all stores level and adapting them at the level of each store.

What if the system was unable to find a comparable promo?

In this case, the coefficient of promotional uplift is predicted by the neural network. The neural network is trained at all promo positions in history, determining the dependence of uplift on every single factor in the promo, as well as on their synergy, and for each product group separately, if there is enough data to take into account the individual characteristics of the items uplift in each group.

How to understand why the neural network predicted specific uplift?

The system has a visualization of the learning results of a neural network, something like a tomography. You can see it in the menu "More - Promo - Correlation of the neural network." There, you will see vertically the factors that NN takes into account when predicting promotional uplift, and you will also see the correlation in percentage of the change in the factor values with the change in the values of the promotional uplift coefficient predicted by NN to the forecast that takes into account or does not take into account price elasticity.

How to organize the process of verification of promotional forecasts and why is it needed?

Promotional sales are the most difficult area in forecasting, as they are influenced by many factors, some of which are not known to the system. For example, in the case of presence or absence of response from competitors, promotions for top-end products, that are traffic generators, can give different uplifts (in general or in different market conditions). Therefore it is important to verify promo forecasts. Using MySales, the forecasting of promos is usually carried out by one person, or a separate group that creates promos in the system, builds an automatic MySales promo forecast and verifies it by comparing it with comparable promos. As well as doing more detailed analysis using the form " Analysis". After such verification of promotional forecasts, it is recommended to upload them to Excel to send to category managers to check for market conditions. If there are substantiated objections from category managers, it is recommended to create a master forecast in the system.

How to take into account promotional sales in the early days of a promotional campaign in order to influence orders during a promotional campaign?

Monitoring the first days of promotional sales is very important for promotions that are valid for a period of time longer than the delivery cycle, so that during the promotion you can respond to the sales and order more or less goods in the nearest delivery. However, if you look at the level of specific positions in specific stores, there may be local extremes or local events that are unlikely to recur in the following days, if we are not talking about top positions. Therefore, in the first days of the promo, it is worth analyzing sales at more integrated levels (SKU-all stores) and adjusting forecast with a Master to reflect the dynamics of the promo for orders during its implementation.

There are positions that grow significantly in the first days of the promo, after which the effect of saturation is observed and sales of subsequent days fall. How to work with this in the system?

If the promotion lasts several weeks and increased sales growth is observed only in the first week, it is recommended to use either the "Additional load" functionality or the auto-order functionality to make an increased order for the first week of the promotion.

How to indicate the percentage of discounts for promos with type MMK?

The discount for such promotions should be indicated as the average percentage of benefits for the buyer, which is provided for a particular position. It is necessary for price elasticity to work correctly in predicting promos.

• If, for example, a promo mechanic has 1 + 1 = 3, then the discount should be indicated as 33% (1 out of 3 units is free)

• For the mechanics “buy 1 and get a discount on the second product”, the discount must be indicated as average for both products. For example, if it is proposed to buy a product for 100 and get a 50% discount on a product for 60, then the total benefit for the buyer will be 30 out of 160, 30/160 * 100% = 19%

• For progressive mechanics, when the more units you buy, the greater the discount, it is better to indicate the average percentage of the discount by analyzing it in the history of such promos. In general, this percentage should be closer to the lowest discount offered as part of the promotion.

There are positions that grow significantly in the first days of the promo, after which the effect of saturation is observed and sales of subsequent days fall. How to work with this in the system?

If the promotion lasts several weeks and increased sales growth is observed only in the first week, it is recommended to use either the "Additional load" functionality or the auto-order functionality to make an increased order for the first week of the promotion.

How to indicate the percentage of discounts for promos with type MMK?

The discount for such promotions should be indicated as the average percentage of benefits for the buyer, which is provided for a particular position. It is necessary for price elasticity to work correctly in predicting promos.

• If, for example, a promo mechanic has 1 + 1 = 3, then the discount should be indicated as 33% (1 out of 3 units is free)

• For the mechanics “buy 1 and get a discount on the second product”, the discount must be indicated as average for both products. For example, if it is proposed to buy a product for 100 and get a 50% discount on a product for 60, then the total benefit for the buyer will be 30 out of 160, 30/160 * 100% = 19%

• For progressive mechanics, when the more units you buy, the greater the discount, it is better to indicate the average percentage of the discount by analyzing it in the history of such promos. In general, this percentage should be closer to the lowest discount offered as part of the promotion.

What if a discount on the product is not provided for the buyer, and instead bonuses are awarded?

In this case, the discount is also better indicated by analyzing the history, but in general, it should be between 1% and percent, calculated as the cash equivalent of bonuses divided by the price of the goods for which they are charged.

How to measure what will be the effect of introducing a promotional module for our network?

There is an indicator of the accuracy of forecasting promos, which is measured as the percentage of promo positions that fall into the allowable error range. This indicator allows you to find the best balance between the level of availability within the promo and the stocks at the exit from the promo. In order to objectively evaluate this indicator, it is important to compare not just the forecast with sales, but also forecast calculated by the system with the forecast prepared by managers manually. If the system predicts better for most positions than managers themselves, then there will be an effect either in increasing the availability of goods in the promotional periods, or in optimizing the stocks at the exit from the promo. It is worth noting that the implementation of the verification process of the system forecasts allows you to get synergy from the use of promotional forecasting algorithms and the expertise of managers.

For what period do you need to prepare and upload a promo history in order to train the system to forecast a promo?

A high-quality forecast of a promo requires at least one year in the promo history, but in some cases two or more years may be necessary. The larger the promo history, the more high-quality promo forecast can be obtained

What if I get a new type of promo for the first time, and there is no history for it? How will the system build a forecast?

Most likely, in this case, the system will forecast this type of promo based on other types of promo averaging forecast. The system will also take into account the characteristics of the products and how they behave in the promo, adapting the promo uplift at the store level based on individual sensitivity. However, with the accumulation of history, the promo system will already begin to take it into account.

How does MySales divides price elasticity and promotional uplift in the forecast?

Price elasticity and promotional uplift are two separate but interconnected components in the forecast. Separate because, first, price elasticity is considered in the basic forecast for those positions where there is enough history. Then, a promotion coefficient is added to the basic forecast, which is adapted at the store level. However, the coefficient of promotional uplift can be added to the forecast, in which there is no price elasticity for those positions where the history is not enough to calculate it. Therefore, the system uses two separate promo uplift coefficients, which are added to the forecast including price elasticity (as a rule, this coefficient is lower) and not including (as a rule, this coefficient is higher). After calculating the promotion uplift in pieces, the engine rebalances these 2 components together with the other forecast components so that they best describe past sales.