What makes time series forecasting hard
Forecasting future values from historical data sounds simple until you try it. The difficulty isn't the math — it's that real-world time series are a mix of three different signals layered on top of each other, and a forecast is only as good as its ability to tell them apart.
The three components
Trend is the long-term direction — revenue generally growing year over year, for example. Seasonality is a repeating pattern tied to a fixed period — weekend traffic spikes, December retail surges, Monday-morning support tickets. Noise (or residual) is what's left over: genuine randomness that no model should try to predict, because it isn't predictable by definition.
A forecast that doesn't separate these will do something recognizably wrong: a model that treats last December's spike as "the trend" will predict a high and rising baseline straight through next summer.
Why naive extrapolation fails
The simplest possible forecast — "next week will look like this week" — fails the moment seasonality exists, because it has no concept of a repeating cycle. A slightly better naive approach — "next week will look like the same week last year" — ignores trend entirely and will systematically under- or over-predict a growing or shrinking business.
Stationarity, and why models care about it
Many classical forecasting methods (ARIMA among them) assume the statistical properties of the series — its mean, variance — don't change over time, a property called stationarity. Most real data isn't stationary on its own (a growing business has a changing mean by definition), so these methods typically work on a transformed version of the data — differenced to remove trend, for instance — and then translate the forecast back.
More recent approaches (Prophet, and sequence models like LSTMs and GRUs) relax this requirement and can model trend and seasonality more directly, at the cost of needing more data and more compute to train well.
What actually determines forecast quality
The honest answer is: how much of the variation in the data is trend and seasonality (predictable) versus noise (not). A sales forecast with strong, consistent seasonality can be genuinely accurate months out. A forecast for something driven by one-off external events has a hard ceiling on accuracy no amount of modeling sophistication will raise — and a good forecasting process says so explicitly instead of presenting a confident number with no uncertainty attached.