Variable Time Windows

dw_volume

I have watched the examples for time windowing, but my data is not always equally spaced, Here is my problem: 

Inventory is done twice a month, but that data is not collocated and made available till ~10 days (8-12 days) after each closing period. The store is only open Monday through Friday and closed on some holidays. During each inventory period, prices fluctuate with some regularity coming on and off sales promotions and waxing and waning demand from the public. We also know how unsatisfied customers come in every day to return or exchange the products across the country, but we don’t know which. Only after the two week inventory periods do we know the return exchange percentage. Returns/exchanges are significant and in poor economic times returns can be extreme as customers want their cash back to buy other items they need, instead of want.  The total of sales and returning/exchanging (but not the mix) customers is known each day. We can rule of thumb that higher than normal returning/exchanging percentages favor returns over exchanges. Exchanges seem to be more a linearly correlated with sales numbers.

In what ways can machine learning forecast from previous twice monthly inventory checks, how many returns or exchanges there were with the daily data. Returns or exchanges only come back in sellable condition, so we don’t need to account for inventory level changes from loss.  There are several hundred products like this and each needs to be calculated separately as the some are more aspirational than necessary for consumers.

Is there a way to handle this in Rapid Miner?

Telcontar120

You could generate time series data (using the Create Exampleset operator) at the most granular level needed (the minimum time interval already present in your data), and then simply join in your inventory records to that set.  Then use Replace Missing Values from the Series operators to fill in the gaps with the prior values, or you could even use linear interpolation to fill in the missings if that works better.

sgenzer

hi @dw_volume - can you please help me understand where this is coming from? It looks like it comes from a textbook. Do you have a data set?

Thanks.

Scott

dw_volume

Yes I have a data set as a rather large SQL Server Database. It has a few items changed to protect my idea but the allegory is correct. The core of the question is about having non-standard time sequence length and spacing of observations. Some think that Dynamic time warping is the way to go. Some suggest there are packages that can do this better.  https://github.com/bnaul/cesium is one example for astronomy where you can't always get series data to come in quantized chunks that simple time sequence methods like.  I could scale the data and post a sample if it helps.

sgenzer

thx @Telcontar120 as usual

@dw_volume feel free to post XML and some public, sampled-down version of your data if you want us to play with it.

Scott