I’ve been working on a forecasting project for a while now, and I keep hitting a wall when it comes to deciding which time series decomposition method is actually appropriate for my data—it’s got some really stubborn seasonal spikes and a trend that seems to shift. I guess I’m just wondering how others have navigated that feeling of being stuck between textbook examples and their own messy real-world numbers.
Yeah I get that feeling the data refuses to behave like the textbook and those seasonal spikes keep moving. time series decomposition can help but it never fully solves the mess. it feels like you are balancing interpretability with fit and you are not alone.
Try a plot followed by a robust decomposition such as STL to see if seasonality changes over time. inspect the residuals for hidden structure. if the seasonal pattern shifts you may need models that tolerate changing seasonality or allow regime shifts.
Maybe the issue is data quality not the method. a few outliers in the spikes can throw off a decomposition. consider cleaning or capping extreme values before testing methods and see if you get a more stable baseline.
Maybe the frame of forcing a single decomposition is the problem. What if you test a regime switching or non stationary model instead and let seasonality arise from the model itself?
I am skeptical that any one method will fit a messy real world series. decomposition sounds tidy but real data often breaks its assumptions and a flexible toolbox beats chasing the perfect frame.
Focus on craft and reader expectations. if this forecast is for a narrative maybe describe regime shifts and breakpoints instead of a smooth trend. time series decomposition may be part of the tale but not the whole plot.