Notes on working with AI
A forecast is a range, not a number
Pre-trained forecasting models can predict your data without training on it. The value is in the error band and the backtest, not in the line.
A forecast without an error band is a guess with a chart
A new kind of forecasting model has appeared: pre-trained on a very large number of time series, then pointed at yours with no training of your own. Google Research's TimesFM is the best-known example. That is real progress, because until recently a usable forecast meant someone fitting and tuning a model per series.
But the line on the chart is the least interesting output. What makes a forecast usable is the range around it and the proof that it has worked before.
What a pre-trained model actually knows
Such a model has seen monthly sales, web traffic, energy demand and thousands of other series. From them it learned general shapes: weekly and yearly seasonality, trends that flatten, the way noise behaves. When you hand it your series, it recognizes which of those shapes your data resembles and extends it forward.
That is also its limit. It knows how series tend to behave, not what happens in your company. A promotion, a strike, a holiday that moved, a new product line: none of it is in the numbers yet, so none of it is in the forecast. Google's own documentation describes the results as comparable to conventional statistical methods. I read that as an accurate sentence: a good baseline that costs almost nothing, not an oracle.
The range is the answer to how sure are you
These models can return a median forecast plus quantiles, for example the value below which the outcome should fall 10% of the time and the value above which it should fall 10% of the time. Together they form a band. A narrow band says the history was regular. A wide band says it wasn't, and you should plan for more than one outcome.
Plan on the band. Staff and cash for the high case, order stock between likely and low, and make the decision rule explicit before the month arrives.
The backtest is the only evidence you have
Before trusting any forecast, hide the last few months, forecast them as if you didn't know, and compare. This takes minutes and tells you whether the typical miss is 3% or 30% on your data. Without it, a confident chart is just formatting. You need roughly two years of monthly history for the test to mean anything, since the model has to see a full seasonal cycle first.
What this is not
It is not a way to predict stock prices. Seasonal sales data repeats because customers and calendars repeat. A market reprices on news that nobody has yet, and a pattern that anyone could extract gets traded away. Nothing here is investment advice.
If you want the model itself, Google publishes it as open source, and the latest weights are licensed for non-commercial use only. For business use, Google offers it through BigQuery. Most readers don't need either: Claude can build a baseline for you from a spreadsheet.
What to do tomorrow
In claude.ai, turn on Code execution and file creation under Settings > Capabilities, upload a monthly sales or expense file, and paste this:
Build a baseline forecast for the next three months from the data I uploaded. For each month give three values: low, likely and high. First hide the last three months, forecast them, and show me in a table how far off you were. At the end, list the business events you cannot know about that could break this forecast.
Look at the backtest error first. It tells you how much weight the range deserves.