[Icathian]

This rings very true to me. I'm working on moving over to an SWE role in the next few years for many of these reasons.

I'll just add one: the business absolutely doesn't care how you get your answer, only if they're reliable enough (hand grenade close is better than most companies have today).

While this seems obvious enough to anyone with a few years under their belt, to the new DS grad who has their time series analysis canned in favor of slapping a simple moving average in place and shipping it can be rather disillusioning.

[proverbialbunny]

Usually, but I've seen the opposite too.

Sometimes a young startup wants to advertise to the board, and they want you to make a presentation. I've made the mistake of showing near 100% accuracy solving a difficult problem important to the business, and expecting a strong positive result.

Instead I got a, "But, are we using deep neural networks?" type comments.

Sometimes a company just wants to market, be it to customers, or to the board. It's important to know your audience.

[apohn]

> the business absolutely doesn't care how you get your answer, only if they're reliable enough (hand grenade close is better than most companies have today).

One of the challenges with this is that "reliable" can mean a lot of things when the goalposts of success are constantly moving in large projects with many stakeholders, all of which are clawing for attention. I've seen politics derail so many Data Science projects and destroy the morale of Data Scientists.

It's only natural that a lot of people will realize that a moving average that confirms what people wanted to see anyway will lead to more success (whatever that means).

[Der_Einzige]

Nothing reliably consistently beats ARIMA models in time series forecasting to this day

That's pretty sad when you think about it, but it's painfully true.

[wenc]

> Nothing reliably consistently beats ARIMA models in time series forecasting to this day

Not sure this is true in practice. In some situations, Holt-Winters (ie. algorithms in the ETS family) may do better, and it's often a good idea to try both.

There's a claim that Holt-Winters is a special case of ARIMA (the claim is ARIMA is more general), but this is actually not the case. There is equivalence in only a subset of cases. [1]

I've fitted Holt-Winters models that beat ARIMA models. ARIMA models can have trouble generalizing from training data with long horizons because they tend to overfit to the distant past. Holt-Winters on the other hand has a natural "forgetting factor" built-in which moderates this.

As well, my experience is that stacked models with well-chosen exogeneous variables (if you have causal variables) tend to outperform pure time-series methods because they are anchored on more independent variables than just t. Pure time-series models bank on the assumption that patterns have a repeatable time-dependence, and most of the time this is just not true, so have to be augmented with other variables.

[1] https://otexts.com/fpp2/arima-ets.html