Point of sale (POS) data can unearth enormous information about the retailer’s lost and hidden opportunities. Since POS data is available over a period of time, the root causes for retailer’s difficulties can be traced easily. The one big challenge with POS is that it is not equidistant; it is a set of transactions. Transactions happen at random intervals; this violates the first assumption of most of time series modeling – evenly spaced data points.
What are the ways to tackle such an issue?
There are many ways we can handle transactions. One of them is to accumulate POS into evenly spaced data. Accumulating data sounds simple but, the challenge is to accumulate in such a way that we do not miss out on any important information regarding the SKU. Misinterpretation of the accumulation interval to be chosen could lead to loss of information which in-turn leads to loss of revenue due to Out-of-Stock (bad inventory planning and hence inaccurate store orders).
For example, Figure 1 shows the point of sale data for yogurt, a fast moving good, for one day. If we accumulate it in intervals of five minutes (Figure 2) we cannot comprehend any distinct selling pattern. Similarly, if we accumulate it in intervals of four hours (Figure 3) then, we miss out on all the selling patterns that happen within the four hours and hence not be able to predict and forecast the actual selling patterns which lead to huge losses and inventory inaccuracies. The optimal accumulation time for yogurt data considered below is 1 hour which gives us a distinct selling pattern represented in Figure 4.
Therefore, it is important to sample each SKU differently as every SKU’s selling pattern defines its own story.
So, how do we sample then?
The right sampling interval for accumulation can be found by studying the velocity of each SKU, through the oscillation period.
The oscillation period for every SKU varies from season to season, the optimal oscillation period is found by calculating the dominant oscillation period, the oscillation period that occurs more often in a given time duration, to determine the accumulation interval. The time so obtained is the optimal interval to accumulate POS.
Once the accumulation is done the assumption of evenly sampled data is satisfied and time-series techniques can be employed on POS to find distinct patterns in the data and hence decipher the tricks of the trade for improving sales by mitigating Out-of-Stock and improvise on revenue by employing the right strategy at the right time.