Choosing the Best Time-Series AI Model for Your Data

Twitter: @mindsdb

[pmiddlet]

Ok so ... not to be a pill here, but pushing a button and having a db centric (looks like it runs on snowflake) makes you look like you developed the forecast model ... so...no. By leaps over sands of the Sahara no :). When someone says 'statistical forecast' - I tend to ask my self -- 'what exactly am I dealing with here?', 'what method(s) it is using or is it a one trick pony?', what was I thinking when I took on this career path in the first place? :)).

It helps to know your data deeply descriptively and with some pre-analysis requirements done first. There are quite a number of numerical and statistical methods (and combos of the 2) that we can apply to the data. But there are 'character traits' of the data that will tell you what methods would even be reasonable, much less meaningful. So IMO, I would never just blindly 'trust the computer'. Know your data, know your methods and sound quantitative principles, and absolutely, know your software.

Also, there are different ways to evaluate the accuracy and precision of models. Blindly thumbing at a table of differences isn't one of them. There are different types of error that lurk within time series. Some of them are about recency, where say, current business operations, or revenue, or whatever process you are describing change the nature of the observed outcomes over time. Linear growth (or whatever outcome of interest you prefer) over time could change to non linear, or heteroscedastic, or demonstrate sharp cusps that move oppositely from the past due to certain significant business events (or an amalgam of a number of things, of which a few are likely to influence the outcome the most). Actual values can often 'drift' from the modeled values, and you have to be able to detect when that is about to happen.

Neural nets for time series analysis isn't likely the best sort of paradigm to use them (especially if you're doing something where the outputs require some explanation to someone, say a reasonably skeptical C-Level. That's not likely with a neural net (black box problem)). ANNs in general appear quite sensitive to more sudden swings in change versus other TS methods. So 'steadier' state with more 'complicated' layers that affect an outcome may be more preferable to an ANNs.

Measures that are often used together to evaluate different 'strengths' and 'weaknesses' of a model - such as MAPE, MASE, RMSE, AIC, BIC. People can spend their entire lives studying time series analysis (some even do it over higher dimensions, like 2 and 3D geotemporal space). For any software that makes this push button - BE A HEALTHY SKEPTIC and heavily test these puppies (and examine their accuracy, precision, AND bias (the leading cause of model drift).