Oracle Retail AI Foundation Cloud Services Implementation Guide Release 22.2.301.0 F59895-02 |
|
![]() Previous |
![]() Next |
This chapter provides details about using Demand Transference.
The DT application assumes that, within a category, all of the items at a store have a common seasonality. This assumption is generally correct for long life cycle categories, where each item does not have a predetermined point of obsolescence or where the point of obsolescence is years from the point of introduction of the item. Examples include most grocery categories and basic clothing items. Electronics items frequently have defined life cycles, generally measured in years.
It is important to address the situation in which different items in the same store have different life cycles and those life cycles are short. In this situation, the store may have items that are at various points in their life cycles and there is no common seasonality. This frequently occurs with fashion merchandise (see "Implementing DT for Fashion Categories").
An assortment elasticity of 0 turns all cannibalization factors into constant 1, meaning the assortment has no cannibalization. This is unlikely. However, it does show that a small-magnitude value for assortment elasticity indicates a category where cannibalization is small. Similarly, a high magnitude of assortment elasticity indicates a category where cannibalization is large. It is possible for the magnitude to be too large.
It is also possible for the Calculation stage of DT to produce assortment elasticities that are positive. Such positive values for assortment elasticity are an indication that there is some unidentified problem with the data, because a positive assortment elasticity means cannibalization factors increase with increasing assortment size, which in turn means each item in the assortment sells more the larger the assortment gets. In the Evaluation stage of DT, such positive assortment elasticities are removed and replaced by assortment elasticities from escalation.
Here is an example that explains the cannibalization model. In this example, all the cannibalization factors are equal; in a real example, the factors would all be different.
Identical, or nearly identical, items are being added to an assortment. If only one of these items is added to the assortment, it takes the entire market share for the item. When another item that is extremely similar is added, the two items split the market share evenly between them. The cannibalization factors are now half for both items. The addition of a third such item creates a three-way even split, one-third for each and the cannibalization factors are each one-third. As more such items are added, the cannibalization factors approach but never reach zero. (Parenthetically, this example also shows how adding items to an assortment does not necessarily produce more market share overall for the assortment, since the new item may simply siphon off sales of existing items.)
The cannibalization factor is actually a power-law, that is, the assortment elasticity enters into the cannibalization factor as an exponent. The cannibalization factor consists of a positive value, the Total Assortment Effect (TAE), raised to the assortment elasticity. Each item in an assortment has its own TAE; the TAE increases as items are added to the assortment. Therefore, the assortment elasticity is a negative number, in order for the cannibalization factor to decrease as TAE increases. (In the example, the TAE could simply be the count of the number of items added so far, and the assortment elasticity would then be -1, thus producing 1/2, 1/3,....) Conversely, the TAE also decreases as items are removed from the Assortment.
Note the similarity to the more conventional idea of power-law price elasticity, which involves a price raised to a negative power (the negative power being the price elasticity). In the cannibalization model, the TAE plays the role of price. Like price, the TAE can change week to week because the assortment can change week to week and the TAE is designed to reflect the current assortment.
The cannibalization factor also accounts for the similarity of the items being added to the assortment, so that similar items cannibalize each other more than non-similar ones. The similarity values are used to calculate the TAE; higher similarities produce a larger TAE, providing a larger decrease in cannibalization factors.
The TAE also depends on the relative sales rates of the items in the assortment. Although higher similarities increase TAE, as discussed above, that cannot be the whole story, because cannibalization depends on the sales rate of the item doing the cannibalizing. For example, consider two items A and B that have equal similarity to an item C that is already in the assortment. Adding item A will cannibalize item C, and so will adding item B; however, if A is a low selling item and B is a high selling item, adding B will cannibalize C more than adding A will. So the TAE must account for the relative sales rates of the items in the assortment in addition to accounting for similarity. The TAE does this through the sales index of an item, which is the sales of the item divided by the average sales of the items in the assortment (total sales divided by the number of items in the assortment). This calculation is done in each week, with the assortment in that week.The following example demonstrates these concepts together and also shows some of the robustness of this model. Consider an assortment that includes items A and B and in which A has a vastly higher sales rate than B. The TAE for item A includes a term for item B, and as discussed above, the contribution of B to the TAE depends on its sales index. In this case, the sales index of B is quite small, because A is so large. Removing B from the assortment will change the TAE of A very little, and so the change in A's sales due to removal of B is quite small. In a situation where A's sales become increasingly large, the sales index of B will go to 0, and the effect of the removal of B on A tends to 0. This is as it should be, since when A becomes so much larger than B, the removal of B cannot have much effect on the sales of A.
The cannibalization factor depends on both the similarity values and the assortment elasticity. It might seem that similarity alone determines cannibalization, as a similarity of 0.5 between items A and B means that A takes half of B's share if A is added, but that is not the case. In particular, by separating the concepts of TAE and assortment elasticity, the model is more robust; if all of the similarity values are biased lower or higher for some reason, the bias can be accounted for by adjusting the magnitude of the assortment elasticity so that the cannibalization factors are still correct.
In order to calculate assortment elasticity, DT requires historical data that contains assortment changes, because DT examines historical data to determine how much cannibalization occurred when historical TAEs changed. From the relationship between changes in historical TAEs and changes in cannibalization, DT then calculates the assortment elasticity. This is similar to calculations of price elasticity. In order to determine price elasticity from historical data, it is necessary to have price changes in the historical data, and the more changes the better.
Using the example from "Assortment Elasticity", suppose, in the historical data for a particular store S, that the cookies assortment has one fewer SKU in week 10 compared to week 1 (that is, some cookie SKU was removed). The TAEs for the other remaining cookie SKUs will all decrease between week 1 and week 10 because of the removal of the one SKU. DT then examines the changes in historical sales units of the SKUs in the cookie assortment at S between week 1 and week 10. By relating the changes in the sales units to the changes in TAEs, DT can calculate the assortment elasticity. A larger magnitude elasticity will result if the changes in TAE caused a large increase in sales units; a smaller magnitude elasticity will result if the increases are moderate.
In reality, the comparisons in this historical analysis that DT does are always more complex than in this simple example. It is rare to find a pair of weeks where the assortment change was just removal of a single SKU. Typically, in each pair of weeks, there are many assortment changes, involving both additions and removals, and the changes in TAE are a result of all of those changes. In the end, though, the relationship between the changes in TAE and the changes in sales units are summarized in a single number, the assortment elasticity, across all pairs of weeks. Because this single number summarizes the number of pairs of weeks and SKUs where TAEs changed, it is necessary an average over all the pairs of weeks and SKUs in the historical data and is not tuned to any particular SKU.
If Category Management Planning and Optimization (CMPO) is used to remove a single SKU from an assortment, it is likely that no pair of weeks in sales history exists in which exactly this SKU was removed and only this SKU was removed. For forecasting the results of this removal, CMPO makes an extrapolation from the historical analysis described above and uses the assortment elasticity that is not tuned to this particular situation of removing only this one particular SKU.
One possible approach to the estimation of assortment elasticity is to estimate it jointly using base rates of sale for each SKU-store and seasonality parameters. However, the estimates of base rates of sale or of seasonality are not required, and an estimation technique that will estimate only assortment elasticity is used. Seasonality is taken into account by assuming that the entire category of items has similar seasonality at a given store, and then applying a differencing technique to get rid of seasonality, leaving base demand and assortment elasticity. For a particular SKU X at store S, the estimation uses pairs of weeks as follows: Suppose W1 and W2 are a pair of weeks where the assortment for the category at W2 at S is different from that at W1. The ratio of sales of X at W2 to sales of X at W1 is taken. This ratio is the dependent variable of the regression, and the independent variable is the assortment difference between W1 and W2. As explained above, the effect of the assortment on X is encoded as the TAE of X and this ratio of sales must be explained in terms of the ratio of the TAE of X between W1 and W2. The input to the regression is many such pairs of weeks for each SKU and each store, and the output of the regression is a single assortment elasticity over all the SKUs and stores of the input.The weeks in each pair are reasonably close together, so that the weeks are more likely to represent similar conditions and provide a better estimate. That is, with such weeks, the only change is likely the assortment change, and it is more likely that the difference in sales units is due to the assortment change than to other external factors.Estimation performs a weighted regression, with the observations from lower-selling SKUs weighted less than observations from higher-selling SKUs, because lower-selling SKUs have more volatile sales and thus should be weighted less according to statistical theory concerning linear regression.Note that since the estimation uses ratios of sales units of SKU X, the absolute quantity of the sales units does not matter (except for the weighting in the weighted regression) for the estimation.The removal of seasonality described above requires that the weeks W1 and W2 in a pair share common items in the assortment within a store. For the pair of weeks to be in the regression, the weeks must have at least five items in common, though this is configurable. Category-store combinations in which the store carries very few items from the category may not meet the requirement of five common items, and a value smaller than five may be necessary, depending on the importance of such category-store combinations. Note that requiring fewer common items results in a poorer seasonality correction, and that five is already a very low number. For example, grocery clients for the demand-transference application usually have assortments numbering in the hundreds.The requirement for common items within-store means each store can have its own seasonality, though this requires a number of common items.In the above description of the regression, W1 and W2 were weeks, but in typical use, the regression uses longer time periods, with a default of four weeks. So W1 and W2 in the default configuration are four-week periods, and the regression uses the average sales of SKU X over those periods. This helps smooth out any noise in the sales of X, and also makes the regression more amenable to low-selling SKUs. Note that extreme low sellers, such as items that sell fewer than one unit in four weeks, will likely still present a problem for the regression.
An assortment elasticity of 0 turns all cannibalization factors into constant 1, meaning the assortment has no cannibalization. This is highly unlikely. However, it does show that small-magnitude assortment elasticity indicates a category where cannibalization is small. Likewise, a high magnitude of assortment elasticity indicates a category where cannibalization is large. It is possible for the magnitude to be too large (see "The Substitutable Demand Percentage").
It is also possible for the Calculation Stage of DT to produce assortment elasticities that are positive. Such positive values for assortment elasticity are an indication that there is some unidentified problem with the data, because a positive assortment elasticity means cannibalization factors increase with increasing assortment size, which in turn means each item in the assortment sells more the larger the assortment becomes. This is presumed to be a nonsensical result, and, in the Evaluation Stage of DT, such positive assortment elasticities are removed and replaced by assortment elasticities from escalation (that is, the elasticities are replaced with higher-level ones).
It is possible, with sufficient data analysis, to figure out what problem with the historical data caused the positive assortment elasticity. However, such analysis is difficult to automate, and escalation is used instead.
The substitutable demand percentage, or just substitutable percentage, of an item in an assortment is the fraction of its demand that is retained by the assortment if the item is removed from the assortment. It is a measure of how substitutable the item is. For example, if the substitutable percentage is 100 percent, then removing the item will not decrease the total sales units of the assortment, since all of the demand for the item will transfer to the other items that remain in the assortment. If, on the other hand, it is 50 percent, then the removal of the item from the assortment means that 50 percent of its demand is lost, and 50 percent is retained. The total assortment sales units will decrease if this item were to be removed from the assortment.
The magnitude of the assortment elasticity has a influence on the substitutable percentage. Increasing the magnitude of the assortment elasticity increases the substitutable percentage. DT only calculates the assortment elasticity for the entire category (not per item), so changing the value of the assortment elasticity changes the substitutable percentage for all items in the category all at once.
It is possible for the magnitude of the assortment elasticity to be too large. This is indicated by a substitutable percentage for several of the items in the assortment that is over 100 percent. A few items can have substitutable percentages over 100 percent, because those are probably outliers. If the assortment is large, it is likely that a few such outliers exist. If 10 percent of items in the assortment are over 100 percent, then the results should be examined.
DT provides a tool for examining the substitutable percentage and for decreasing the assortment elasticity if too many items have a substitutable percentage over 100 percent. Here are some guidelines for using this tool.
When selecting the time interval for the tool, select one that is likely to contain assortments that are representative of the retailer's current assortments. Since the retailer is going to be using the assortment elasticity in forecasts of what happens when current assortments are modified, it makes sense to test the assortment elasticity against assortments that are as similar as possible to the current ones.
It is possible to use the tool to dial down the assortment elasticity. Using Setting maximum substitution percentage, DT calculates an assortment elasticity that results in substitution percentages that do not exceed the set maximum. When using this feature, you may want to set the maximum to a value higher than 100 percent if there are some outlier items that have high substitution percentages. Forcing these outliers down to 100 percent may result in a small-magnitude assortment elasticity, which may mean unacceptably small substitution percentages for all except the outlier items. You may want to select a maximum that is higher than 100 percent but that still brings most items down to 100 percent, leaving a few outliers above 100 percent.
It is possible to use this tool to set the maximum percentage even if all substitution percentages are already below 100 percent. You may have business knowledge, or a directive from the retailer, and know that a particular category must exhibit a substation percentage of at most 70 percent. In this case, this tool can be used to bring the substitution percentages down to 70 percent. This can make the difference between acceptance and rejection by the client.
A time interval for the CDT calculation can be set in the CDT Data Setup stage. No equivalent exists in the Data Setup stage of DT.
The cannibalization factor directly incorporates information about the assortment through the TAE, and so the cannibalization model can handle large assortment changes. This makes it less necessary to use a time interval for DT, compared to CDT, because historical assortment changes can be directly accounted for in the model as changes in TAE.
In the Calculation Stage for both DT and CDT, it is possible to set up the calculation so that it is performed at all combinations of levels of the segment hierarchy and the location hierarchy. This is a more practical possibility for assortment elasticity than for the CDT calculation, because the assortment elasticity is not examined directly by people (unlike the CDTs), and producing thousands of values will not cause an issue. However, it is recommended to use only one of the two hierarchies in the Calculation Stage. Set either the segment hierarchy or the location hierarchy (or both) to be Chain. Because the calculation of assortment elasticity requires assortment changes in history, generating assortment elasticities at all levels may mean that, at lower levels, the data does not contain enough assortment changes in history. You may want to use your business knowledge of the particular retailer or particular category here, since you may know for the retailer or for the category whether assortment changes are frequent or not in the historical data you have. If the assortment changes are infrequent, you may be better off calculating a Segment- Chain/Location Chain assortment elasticity only.
The last stage in DT involves setting the escalation path. If you are using only the segment hierarchy or only the location hierarchy, the escalation path is simply the hierarchy that you are using, and you set the escalation path according to the hierarchy. If you are using both a location hierarchy and a segment hierarchy, then usually you should set the escalation path to go up the segment hierarchy first, and then the location hierarchy. It is better to use only one of the hierarchies.
When using both hierarchies, the escalation path is necessary in order to tell the application which parent it should go to when moving up from a given segment/location node. With both hierarchies in play, every segment/location node has multiple higher-level nodes that do not lie along a single path. The escalation path is necessary to tell the application in what order the higher-level nodes should be considered. When only one hierarchy is used, the higher-level nodes form a single path.
DT can automatically and periodically update the assortment elasticities as new sales history is available. This feature is unique to DT; CDT does not perform automatic updating because it makes less sense to automatically produce new CDTs. New assortment elasticities can be loaded into the consuming applications and thus immediately used; however, the value of new CDTs is less clear.
When new historical transactions enter the RA schema, DT will automatically aggregate them and produce new SKU-store-week sales-units aggregates. These new aggregates are then appended to the older SKU-store-week aggregates, and the resulting data set is then used in a new calculation of assortment elasticities.
Note the following about this calculation:
It does not in any way run the full DT application, that is, re-run all of the stages. The calculation is more targeted and just calculates assortment elasticity.
It only updates assortment elasticities, not the similarities from the Similarity Calculation Stage.
Because it uses a mix of old data and more recent data, the values of the assortment elasticities will change slowly over time as the data set becomes more tilted towards newer data. This is by design. It is not desirable to have sudden changes in assortment elasticity, since that would result in sudden changes in cannibalization and demand transference.
Any assortment elasticities that were overridden using the Substitutable Percentage tool (see "The Substitutable Demand Percentage"), stay overridden, and are not updated.
It is possible for a retailer to have categories where the assortments are very small, that is, 20 or fewer items in the assortment. Such categories can pose a problem for DT because of the small amount of sales data for just 20 items, and also the number of assortment changes may be quite few.
It is better if the assortment is small but items from a much larger set have been added or removed frequently from the assortment. That is, the category has a much larger set of items, but only 20 of them are in an assortment at a given time. It is possible that the assortment changes were frequent enough that more than 20 items have sales history, and in this case DT results may be reliable even though the assortment is small.
For various reasons, fashion categories require some special consideration. This section describes what is different about them and how to handle the differences.
The lowest level of data must not be the SKU level, that is, the Size level of the merchandise hierarchy. Because size is a functional-fit attribute, or nearly so, the level of calculation must be at least one level above size (Style-Color). The historical sales-units data must be aggregated at least up to Style-Color. This also helps avoid problems with low sales rates and noisy sales rates at the SKU level, both common problems for fashion categories. It also helps decrease the number of SKUs within a category, since a multitude of sizes are possible for each Style-Color.
It is worth considering whether color is necessary. Aggregating to the Style level means that transference among colors cannot be calculated. However, it is not clear how useful calculating transference among colors would be, since the colors change for every selling season, and calculating historical transferences among colors may not be particularly useful. A possible halfway approach here might be to aggregate to Style-Primary-Color, where there are only a few primary colors. The primary colors chosen can be the ones that are stable season after season, so that historical transferences among them might be useful in future selling seasons. The primary colors can be chosen to be groupings of the actual colors (so for example, midnight blue and sky blue would become blue). In general, in fashion, the number of colors can be large, and it is unlikely that calculating transferences among such colors would be useful. Aggregating to Style-Primary-Color or even to Style can help avoid low sales rates and noisy sales rates.
It is possible to employ different approaches in grouping the colors. One is to use the primary color. It is also possible to group the colors based on the type of customer the color is designed to attract. For example, the colors can be grouped in "trendy colors" vs. "basic colors." The grouping should be decided in consultation with the retailer, to determine how the retailer uses colors in the category. The retailer may already have a grouping of colors that it uses, and the simplest approach may be to use this grouping.
Here, it is assumed that the aggregation of the historical sales-units data to be either Style-Primary-Color or Style. The term "item" should be understood to mean Style-Primary-Color or Style, depending on the chosen aggregation.
Typically, for fashion, the number of colors is large because of all the color variants. If it is necessary to retain all of the colors instead of following the recommendations above, then it will be necessary to split the color attribute into at least two attributes, a primary color and a secondary color. For more information, see "The Role of Attributes in Calculating Similarities".
DT makes the assumption that the items within a category at a store all have a common seasonality (see "Seasonality in Historical Sales Data"). Because of the short (tens of weeks) life cycles of fashion items, and because items within a category may have different introduction times within the same store, the assumption of common seasonality across the items in a category is probably not valid for fashion categories. It is possible within the category to have items that are at various points in their life cycles within the same store. At the same time, some items may be in the uptrend part of their life cycle, while other items are in the downtrend part of their life cycle.
There are some ways to deal with this by properly setting up the input data for DT.
One simple approach is to approximate the life cycle of an item by using the SKU-store-ranging described in Demand Transference. In this approximation, the range for an item is set to start at x percent sell-through of the item and end at y percent sell-through. Choose x to be 5, and y to be 70. X must not be 0 (that is, the point of introduction of the item), since it takes sometime after the point of introduction of the item for customers to start buying it in quantity and for the item to start having any kind of cannibalization effect on the other items. Y must be set to a point where significant numbers of customers have started to either lose interest in the item or where the item no longer has sufficient numbers of sizes available. In either case, customers are now transferring their demand to the other items in the assortment, so it is as if the item were no longer in the assortment. This is an approximation as the item is still in the assortment and is still selling, just at a significantly lower rate than its peak sales rate.