1 of 34

The Inventory Day

or how not to win the MELI Data Challenge 2021

Seminario virtual a cargo del Lic. Pedro Pury

Organizado por el grupo de investigación sobre Análisis y Procesamiento de Grandes Redes Sociales y Semánticas

Noviembre 30, 2022

2 of 34

pedro.pury@unc.edu.ar 22/11/30

Summary

The Challenge
The Data
The Metric
The Baseline
The Benchmark
Gaussian Model (ML)
Stochastic Model
Evaluation
Concluding Remarks

3 of 34

pedro.pury@unc.edu.ar 22/11/30

0. Disclaimer

I do not work nor did I work at Mercado Libre
I didn't participate at the MeLi Data Challenge 2021
All the data used in this talk was made public on the official website of the event
This is not a talk about ML but I will make some comments on ML.
Opinions are mine alone and do not necessarily represent my institution.
I'm presenting a work in progress.

This talk is just another episode of the human-machine war...

4 of 34

pedro.pury@unc.edu.ar 22/11/30

1. The Challenge

Inventory problem

Papers since 50's, even a book in 2021

Inventory day (iday)

The task is to predict how long it will take for the inventory of a certain item to be sold completely. In inventory management theory this concept is known as inventory days.

SKU: Stock Keeping Unit

5 of 34

6 of 34

7 of 34

pedro.pury@unc.edu.ar 22/11/30

8 of 34

pedro.pury@unc.edu.ar 22/11/30

2. The DATA

Ml-challenge.mercadolibre.com → HTTP ERROR 503 : (

3 sets

Training set for 660916 SKUs

2 months (February and March 2021) 59 days

> 38.9 x 10⁶entries

For do predictions on April 2021 30 days

9 of 34

10 of 34

11 of 34

12 of 34

13 of 34

14 of 34

pedro.pury@unc.edu.ar 22/11/30

3. The Metric

15 of 34

pedro.pury@unc.edu.ar 22/11/30

Point estimation

RPS = |u₀- u|

16 of 34

pedro.pury@unc.edu.ar 22/11/30

4. The Baseline the uniform distribution

17 of 34

pedro.pury@unc.edu.ar 22/11/30

4. The Baseline the uniform distribution

<RPS> = d/6

18 of 34

pedro.pury@unc.edu.ar 22/11/30

5. The Benchmark

19 of 34

pedro.pury@unc.edu.ar 22/11/30

6. Gaussian model

20 of 34

pedro.pury@unc.edu.ar 22/11/30

6. Gaussian model

21 of 34

pedro.pury@unc.edu.ar 22/11/30

7. Stochastic model

22 of 34

pedro.pury@unc.edu.ar 22/11/30

7. Stochastic model

23 of 34

pedro.pury@unc.edu.ar 22/11/30

7. Stochastic model

24 of 34

pedro.pury@unc.edu.ar 22/11/30

7. Stochastic model

25 of 34

pedro.pury@unc.edu.ar 22/11/30

26 of 34

pedro.pury@unc.edu.ar 22/11/30

27 of 34

pedro.pury@unc.edu.ar 22/11/30

8. Evaluation

Example: SKU 538100

Sales in 28 days of february 2021

0 0 2 1 2 0 0 0 0 1 0 2 1 0 0 0 0 0 0 1 0 0 2 1 0 0 1 1

Sales in 31 days of march 2021

0 1 2 0 0 0 1 0 1 0 3 1 0 0 1 0 1 1 0 0 0 0 0 2 0 1 0 1 2 3 4

1 3 4 5 8 9 10 11 12 14 15 16 18 21 25 m

28 of 34

Pedro Pury 22/11/30

29 of 34

Pedro Pury 22/11/30

30 of 34

Pedro Pury 22/11/30

31 of 34

Pedro Pury 22/11/30

32 of 34

Pedro Pury 22/11/30

Rank 12

495353 SKUs

3254422 evaluations

33 of 34

pedro.pury@unc.edu.ar 22/11/30

9. Concluding remarks

The model retrieves the most valuable information from the time series,

when there are data...

Does not require "learning" (training in the ML sense)
The model not only predicts the inventory day, but also gives the daily probability of a given stock. For ML this represents another different training problem
Contributes to the understanding of iday prediction
Given that it does not use contextual information, it cannot predict when

there is no data. It is not good enough to allow you to win the challenge :(

But it could be a tool to improve an ML model: ideas are accepted!!

34 of 34

pedro.pury@unc.edu.ar 22/11/30

That’s all folks !

GRACIAS