-1
$\begingroup$
  • PROJECT: I am working on an e-commerce site where digital products can run out so there is need to reorder them 72h before they run out (reordering them sooner is not a problem but having notification a bit later so if the product would sell better that would be a problem because we cannot reorder products in time).
  • GOAL: is to know if products run out at least 72h earlier.
  • DATA COLUMNS: sales datetime, product id, current number of products, price of product, what currency it was purchased in, other data like profit currency was used for the purchase…
  • SIZE: Before grouping I have a few millions of rows after grouping hundreds of thousands so it is a lot of data point but DASK can handle them.
  • GRUPPING COLUMNS: I have grouped the data by PURCAHSEDATE & ID so each day has the product that were sold with all its feature. Features have been aggregated mostly buy summing (profit, expenses) and mean (percentage features like margin%)
  • HOW FAR I HAVE GONE WITH THE PROECJT: I have looked up a couple of Kaggle projects online that were focused on use https://www.kaggle.com/tejasrinivas/simple-xgb-starter
  • PROBLEMS: A.) Some product has been sold in the past but they are selling out in 1-2 days so it is hard to put trendline on it. B.) Some item just has 1-2 days of data because it just started to sell a few days ago. C.) I also have data of products that have been sold a lot for a mid or long run (hundreds of days thousands of times). So I could do time series modelling on the whole of the sales but for each individual item I don't always have data on it
  • CURRENT RESULTS: I have used XGBOSOT Regression like It predicts well number of products sales after the days is over with all the features, but that is not the goal - https://www.kaggle.com/tejasrinivas/simple-xgb-starter
  • PROJECT RECOMMEND:I am trying to use the following pick ideas from the following competitions: https://www.kaggle.com/c/demand-forecasting-kernels-only/notebooks?competitionId=9999&sortBy=voteCount , https://www.kaggle.com/c/competitive-data-science-predict-future-sales/code
  • GOAL: simple and easy solution, not LSTM or something complicated but something quick and easy (like xgboost regression so if I have more data I can use rapids.ai to GPU teach it) to implement because as I said it is not a problem if it is missing on the time frame on the positive side and the item gets reordered 96h early and not 72h early. I am guessing that somehow, I should shift the dates but as I said in many case items have not enough dates to shift their sales date.
$\endgroup$
0
1
$\begingroup$

If you have sold only once or very few items you will need some prior input (domain knowledge). One term for search is intermittent time series. Here is a stored search.

When you have many time series, related, and interest in both totals and single series, that is called hierarchical forecasting. One expert is here (the author of that blog was the founder of sister site Cross Validated).

With time series forecasting it is often difficult to beat simple methods, see https://stats.stackexchange.com/questions/135061/best-method-for-short-time-series/135146#135146

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.