Smarter Spending Predictions with Machine Learning

When your budgeting app tells you "you're on track to spend $3,200 this month," it's usually just dividing your spending so far by the number of days passed. That works on day 28. It's useless on day 5. Clarity uses Google's TimesFM 2.5 foundation model to generate spending predictions that actually account for how you spend — not just how much.

Why Linear Extrapolation Fails

Linear extrapolation treats every day as identical. Spend $500 in the first 5 days, and it projects $3,000 for the month. But those first 5 days included your $1,200 rent payment. The remaining 25 days won't look anything like the first 5.

The fundamental problem is that personal spending is non-stationary. It varies by day of week, by proximity to payday, by time of month, and by season. Averaging it into a daily rate destroys all of that information.

Consider a real pattern: someone who spends heavily on weekends (restaurants, entertainment, shopping) and minimally on weekdays (just lunch and gas). Their weekend spending might be 3x their weekday spending. If the month so far has included more weekends than the remaining days, a linear projection will overestimate. If fewer weekends have passed, it will underestimate. Either way, it's wrong in a predictable way that a smarter model can correct.

How TimesFM Learns Your Spending Shape

TimesFM 2.5 is a time-series foundation model from Google Research that was pre-trained on billions of time-series data points across diverse domains. When applied to your spending data, it recognizes temporal structures that are common across all types of sequential data: periodicity, trends, level shifts, and seasonal effects.

Clarity feeds TimesFM your historical daily spending totals — typically 60 to 180 days of data. The model identifies recurring patterns in that sequence and uses them to predict each remaining day of the current month. The key patterns it captures:

Day-of-Week Patterns

Most people have a consistent weekly spending shape. Maybe you spend $35 on a typical Monday (coffee, lunch, transit) and $150 on a typical Saturday (groceries, dinner out, activities). TimesFM learns this 7-day cycle from your history and applies it to the forecast. If the remaining days of the month are Tuesday through Sunday, it predicts each day according to its typical weekday spending level, not a flat average.

Pay-Cycle Effects

People who are paid biweekly or semi-monthly tend to spend more in the days immediately following payday. This creates a roughly 14-day or 15-day cycle in spending data. TimesFM detects this periodicity without being told your pay schedule. It sees the pattern in the data and extrapolates it forward.

This is especially important for the budget pace feature. If you're 3 days past payday and spending is elevated, a linear model flags you as over-budget. The TimesFM model knows this spike is temporary and that spending will taper in 4-5 days. It won't send you a false alarm.

Monthly Recurring Spikes

Rent on the 1st. Car insurance on the 15th. Internet bill on the 22nd. These create predictable spikes at fixed points in the month. TimesFM sees these as sharp, periodic increases in the daily spending series and expects them at the same points in future months. A linear model, by contrast, either over-counts them (if they've already happened) or ignores them (if they haven't).

Seasonal Trends

With 3+ months of history, TimesFM can detect seasonal spending shifts. Higher spending in December (holidays), lower in January (post-holiday belt-tightening), higher in summer (travel, activities). These month-to-month shifts inform the model's baseline expectations for the current month.

Confidence Bands: A Range, Not a Number

A point prediction ("you'll spend $4,100") is always wrong by some amount. The useful question is: how wrong? TimesFM produces quantile predictions that give you a calibrated range.

Clarity shows three levels on the dashboard spending projection:

• Likely range (p25-p75): You have a 50% chance of landing in this range. Example: $3,900 - $4,300.
• Extended range (p10-p90): You have an 80% chance of landing in this range. Example: $3,600 - $4,600.
• Central estimate (p50): The median prediction. Example: $4,100.

The practical benefit: instead of "you'll spend $4,100," you see "you'll likely spend between $3,900 and $4,300, and almost certainly between $3,600 and $4,600." That's actionable information. If your savings goal requires keeping spending under $4,000, and the model says there's only a 25% chance of that, you know you need to make active changes.

Where Predictions Appear in Clarity

The TimesFM spending predictions aren't buried in a separate analytics page. They're woven into the features you already use:

• Dashboard spending card. The month-end projection on your main dashboard uses the ML prediction. The confidence band appears as a shaded area on the spending trend chart, widening as the forecast extends further into the future.
• Budget pace indicators. Each budget category shows whether you're ahead or behind pace. That pace calculation now uses predicted spending patterns rather than a linear daily rate. Your grocery budget won't show "over pace" just because you did a weekly stock-up on Saturday.
• Spending insights. The insights bar on the dashboard surfaces observations like "your projected spending this month is 12% higher than your 3-month average" — using the forecast, not a naive projection.
• AI assistant. Ask "will I stay within budget this month?" and the assistant uses the TimesFM forecast to give you a probability-weighted answer, not a yes/no based on a straight line.

Accuracy vs. Linear: What We See

The difference between TimesFM and linear predictions is most pronounced in the first half of the month. By day 5, a linear projection can be off by 40-80% because it's extrapolating from a small, unrepresentative sample. TimesFM forecasts at day 5 are typically within 10-15% of the actual month-end total because they use pattern information from prior months, not just the current month's data.

The gap narrows as the month progresses. By day 20, both methods are reasonably close because 2/3 of the actual data is in. But the first two weeks are when forecasts matter most — that's when you have the most time to adjust your behavior.

What the Model Doesn't Do

TimesFM predicts aggregate daily spending totals. It does not:

• Predict individual transactions or merchant-level spending (it works on daily totals)
• Account for one-time events it hasn't seen before (an unexpected car repair won't appear in the forecast until it happens)
• Replace your judgment about upcoming known expenses — if you know you have a $500 medical bill coming, the model doesn't know that unless it's a recurring pattern

It's a pattern-based forecast, not a crystal ball. Its value is in capturing the predictable structure in your spending that linear models ignore. For most people, that predictable structure accounts for 70-80% of their monthly spending variance.

Getting Better Predictions

The quality of TimesFM predictions improves with more and cleaner data:

• Connect all your spending accounts. If half your spending goes through a credit card that isn't connected, the model only sees half the pattern. Connect checking accounts, credit cards, and any other accounts you spend from.
• Give it time. Predictions with 30 days of history are decent. Predictions with 90+ days are significantly better. The model needs enough history to distinguish a real pattern from noise.
• Don't worry about categorization. TimesFM works on raw daily spending totals, not categories. You don't need perfect categorization for accurate forecasts.

The goal isn't to predict your spending to the dollar. It's to give you a realistic range early enough in the month that you can act on it. A forecast that says "you'll spend between $3,800 and $4,400" on day 7 is infinitely more useful than one that says "$6,200" because your rent hit on day 1.