,
9 tweets,
2 min read
Read on Twitter
Forecasting is often made more complex than it needs to be. Given your alternatives are guessing (or hoping), even a simple model will perform better. Here is my thinking process 1/
My first goal is to get a baseline that real (future) data an prove the forecast is reliable. Until then, I hide the most recent periods of data and see how my model would have performed using data from the period before that. 2/
Is the work we are about to do similar in complexity and unknowns as prior work. If it isn't, I IGNORE historical data. Sometimes historical data matches just because the system is more at play than the work itself. 3/
How can I tell if the system or the work complexity dominates? Plot looking at lead times against size buckets. If they are correlated at all (linear pattern), then work size matters, if they don't correlate its system dominated 4/
Another (geeky) way is I look at the work cycle time distribution - if it's exponential it is the work; if its bell curve peak shifted left, its system dominated. 5/
If its system dominated, I ignore work size estimates and forecast what the system throughput rate is likely to be and what constraints limit that (normally a skillset). Split rate is important here, work grows at a multiplier rate (normally 1 to 3x). 6/
If its work dominated, I look at work size in particular items that might blow out into extremely large size (things never attempted before, things nobody yet knows how to size). 7/
In all cases, I look at RISKS. All too often "most" of the work is complete but delivery blocked due to one outstanding item or defect. My forecasting focuses on finding what item that might be. 8/
Once real data is flowing (throughput rate, defect rate, work splitting rate) I adjust the model to forecast well. AGAIN I only belive my future forecast if it proved reliable in recent history. 9.
