How AI Cash Flow Forecasting Works in Clarity

Most finance apps estimate your future cash flow by dividing what you've spent so far by the number of days elapsed. That gives you a flat line that ignores everything about how you actually spend money. Clarity replaces that with a machine learning model that understands your weekly rhythms, seasonal patterns, and pay-cycle effects — powered by Google's TimesFM 2.5 foundation model.

The Problem with Linear Cash Flow Projections

Open any budgeting app on the 15th of the month. You've spent $2,100 so far. The app divides $2,100 by 15 days, gets $140/day, and projects you'll spend $4,340 by month end. Simple math. Completely wrong.

Here's why that fails: your spending isn't uniform. You spend more on weekends than weekdays. You spend more in the first few days after payday. You have recurring bills that hit on specific dates. The last week of the month might look nothing like the first week. A flat daily rate smooths all of that away and gives you a number that's consistently off by 15-25%.

If you're paid biweekly and your rent hits on the 1st, your spending in the first three days of the month might be $1,800 (rent + groceries + gas). A linear model sees that and projects you'll spend $18,600 for the month. That's not a forecast; it's a panic attack.

How TimesFM Changes the Calculation

Clarity uses TimesFM 2.5, a time-series foundation model developed by Google Research, to generate cash flow forecasts. Instead of drawing a straight line from your spending so far, it analyzes your historical daily spending patterns and predicts what each remaining day of the month will actually look like.

TimesFM is a foundation model for time series, meaning it was pre-trained on billions of real-world time-series data points before being applied to your financial data. It understands temporal patterns out of the box — seasonality, trends, cyclical behavior — without needing months of your data to learn from scratch.

When Clarity generates a cash flow forecast, it feeds your historical daily spending totals into TimesFM and asks it to predict each remaining day in the month. The model produces not just a single point estimate but a full probability distribution for each day, expressed as quantile predictions.

What the Model Captures

TimesFM picks up on patterns that a linear projection completely ignores:

• Day-of-week effects. You spend $45 on average on Tuesdays but $120 on Saturdays. The model learns this rhythm and applies it to the remaining days in the forecast window. If there are two Saturdays left in the month, it accounts for two high-spending days, not two average ones.
• Pay-cycle patterns. If you're paid biweekly, your spending tends to spike in the 2-3 days after each paycheck. The model recognizes this cycle and adjusts its predictions accordingly. It won't assume you'll spend the same amount on day 20 (mid-cycle) as day 1 (just paid).
• Recurring bill timing. Your $150 car insurance hits on the 22nd every month. Your $65 internet bill hits on the 5th. The model sees these consistent spikes in your history and expects them at the same times going forward.
• Seasonal trends. You spend more in December than February. More in summer than spring. TimesFM captures these longer-range patterns across months of history, so your December forecast doesn't look like your February forecast.
• Trend shifts. If your overall spending has been gradually increasing (or decreasing) over the past few months, the model incorporates that directional trend rather than assuming stationarity.

Confidence Intervals: The p10/p90 Bands

A single number ("you'll spend $4,200 this month") is almost certainly wrong. The question is how wrong. That's why Clarity shows confidence bands alongside the central forecast.

TimesFM produces quantile predictions at multiple levels. Clarity uses three:

• p50 (median): The model's best single estimate. You have roughly a 50% chance of spending more than this and 50% chance of spending less.
• p10 (optimistic): There's only a 10% chance you'll spend less than this amount. This is your "best realistic case."
• p90 (conservative): There's only a 10% chance you'll spend more than this. This is your "worst realistic case."

On the Clarity dashboard, these show up as a shaded band around the forecast line. The band is narrow when the model is confident (consistent historical patterns) and wide when there's more uncertainty (volatile spending history, limited data).

For example, if your forecast shows "$3,800 - $4,400 (likely $4,100)", that means the model expects you'll most likely spend around $4,100, but you should plan for anywhere between $3,800 and $4,400. That range is far more useful than a single number that pretends to be precise.

Before and After: A Practical Example

Consider someone who earns $5,000 biweekly, with paychecks on the 1st and 15th. Their monthly spending is typically around $4,200, but it's front-loaded after each paycheck.

Linear Projection (Before)

On the 10th of the month, they've spent $2,400. Linear math: $2,400 / 10 days = $240/day, projected total = $7,440. That's 77% higher than their actual monthly average. Why? Because the first 10 days included rent ($1,600) and a post-payday grocery run ($180). The linear model doesn't know those won't repeat.

TimesFM Forecast (After)

The model knows that days 11-14 are typically low-spending ($40-60/day), that spending picks up again on the 15th (payday), and that the second half of the month averages about $1,900 in total. Its forecast: $4,300 (p50), with a range of $3,900 - $4,700. That's within 5% of their actual historical average instead of 77% off.

Where This Shows Up in Clarity

The TimesFM cash flow forecast integrates into several places across the app:

• Dashboard spending card. The projected month-end spending total on your dashboard uses the ML forecast instead of linear extrapolation. The confidence band appears as a subtle range beneath the primary number.
• Budget pace warnings. When Clarity warns you that you're on track to exceed a budget category, it uses the forecast model to determine pace. A spike in grocery spending on the 3rd doesn't trigger a false alarm if the model expects your grocery spending to taper off mid-month.
• AI chat responses. When you ask the Clarity assistant "how much will I spend this month?", it returns the TimesFM forecast with the confidence range, not a simple division.

How Much History Does It Need?

TimesFM works best with at least 60-90 days of daily spending data. With less than 30 days, Clarity falls back to simpler heuristic projections. The model improves as it gets more history — 6 months of data produces noticeably tighter confidence bands than 2 months.

Because TimesFM is a pre-trained foundation model, it doesn't need to be retrained on your data. It already understands time-series patterns. Your historical spending is provided as context at inference time, and the model generalizes from its pre-training to your specific patterns. This means you get reasonable forecasts much faster than you would with a model trained from scratch.

Privacy and Processing

The forecasting model runs on aggregated daily totals, not individual transactions. TimesFM sees a sequence of numbers like [142, 87, 210, 45, ...] — it doesn't know or need to know that $142 was spent at Whole Foods versus Target. Your merchant names, account details, and transaction descriptions are never sent to the model.

What This Means for Your Finances

Accurate cash flow forecasting changes how you make decisions. When you know you'll likely spend between $3,800 and $4,400 this month (not a vague "$4,100 projected"), you can:

• Set aside the right amount for savings — using the p90 estimate to be conservative
• Avoid unnecessary anxiety from inflated linear projections early in the month
• Spot months where spending is genuinely trending high versus months where it just looks high because of bill timing
• Make informed decisions about large purchases by seeing how they'd shift the forecast

The gap between linear projections and ML forecasts is largest at the beginning of the month and narrows as more actual data comes in. By the 25th, both methods converge. But the first two weeks — when you're making the most decisions about the rest of the month — is where TimesFM makes the biggest difference.