Math first, narrative second

The forecast starts with arithmetic you could redo by hand. For every category, vokse takes the last six months of activity and smooths them with an exponential moving average (α 0.5), so the most recent month weighs most and older months fade without vanishing. That figure is the baseline, and it is computed in integer cents.

On top of the baseline sit the recurrences you already keep: salary, rent, subscriptions, scheduled transfers. Every occurrence that falls inside the target month is added to its category, and the sum is compared with what you have assigned in the budget, so the overrun is visible before the month begins. The same inputs always give the same numbers, and this path runs without any AI provider or token budget.

The narrative comes second and stays optional. Ask for the refined pass and a language model reads the baseline and the six months behind it, adjusts only the categories where a trend or a season clearly says next month will differ, and attaches a one-line reason to each change. It cannot add a category or a figure the math did not produce; if the call fails or is off, the deterministic forecast stands as it was.