/sci/ - Science & Math

>>12593040
Appreciate the hints and help.

It's a fixed length time series, with a small number of values per day in the series. Also has a ton of other global metrics that apply to the entire time series we're including too.

We're trying to predict one metric from the next day out of sample in the time series. For a concrete example:

We have 365 days of data for 10,000 shops. Each day includes data for:
- the number of toilet paper rolls sold per day
- the daily customers
- the shop turnover
- toothbrushes sold

This gives us a per-day vector of [math]\{ k_1, k_2, k_3... \} [/math]

Our global parameters would be something like:
- size of shop in square meters
- number of cash registers
- hours shop is open per day
- average tp rolls sold
- average daily customers
- std dev of tp rolls sold
- etc.

This gives us a global vector of [math]\{ g_1, g_2, g_3... \} [/math]

All of this data is concatenated into a single input vector, where [math]i[/math] is the day in question.

[math]
\{ k_{i0}, k_{i1}, k_{i2} ... k_{(i+1)0}, k_{(i+1)1}, k_{(i+1)2}, ... g_{1}, g_{2}, g_{4} ...\}
[/math]


We're trying to predict the number of toilet paper rolls sold on the 356th day for a given shop.

Hopefully this help clarify how retarded I am.

>>12593110
Thanks, reading up on this. I'm not sure how PCA would help if we're not seeing any correlation at all? At the moment we're seeing MAPE of like several thousand %.