Anomaly Detection: How Clarity Spots Unusual Spending

You check your bank statement and see a $340 charge you don't recognize. Or your grocery spending quietly doubled over 3 months and you didn't notice. Most finance apps won't catch either of these. Clarity uses the same TimesFM model that powers its forecasting to detect when your actual spending falls outside the range of what's expected — automatically, without you setting manual thresholds.

How Time-Series Anomaly Detection Works

Traditional anomaly detection in banking uses fixed rules: flag any transaction over $500, flag any purchase in a new country, flag any charge at 3am. These rules catch obvious fraud but miss subtle anomalies and generate false positives for anyone whose spending doesn't fit a narrow template.

Clarity takes a different approach. TimesFM 2.5, Google's time-series foundation model, generates a one-step-ahead predictionfor each day based on your spending history. It predicts what today's spending should look like given your patterns, then compares the prediction to what actually happened.

The model doesn't just predict a single number. It produces quantile predictions — a range of expected values at different confidence levels. The p10-p90 band represents the range where your daily spending should fall 80% of the time. When actual spending lands outside this band, something unusual happened.

What Makes This Different from Alerts

Most apps let you set spending alerts: "notify me if I spend more than $200 in restaurants this week." That requires you to know what thresholds to set for every category, and to update them as your spending changes. It's manual and brittle.

TimesFM-based detection is adaptive. The expected range adjusts automatically as your patterns change. If you gradually start spending more on dining out, the model's expected range shifts upward with you. It won't flag your new normal as an anomaly. But if your dining spending suddenly doubles in a single week — that's outside the expected range, and it gets flagged.

This means it catches two types of anomalies that fixed thresholds miss:

• Unusually high spending on a normally low day. You typically spend $40 on Tuesdays. The model expects $25-$65 (p10-p90). If you spend $180 on a Tuesday, it's flagged. A flat "over $200" alert would miss this.
• Unusually low spending when high spending is expected. You always have a $1,600 rent payment on the 1st. If the 1st passes with no large charge, the model notices the absence. This can catch failed autopay or a payment that didn't process.

Use Cases: What Anomaly Detection Catches

Fraudulent or Unauthorized Charges

If someone uses your card for a $340 purchase at a store you've never shopped at, that charge contributes to a daily spending total that's well above the expected range. The anomaly flag prompts you to review the day's transactions and spot the unauthorized charge.

This isn't a replacement for your bank's fraud detection — banks are better at catching the classic fraud patterns (foreign transactions, card-not-present at unusual merchants). But Clarity catches the cases that slip through: charges at legitimate-looking merchants, small repeated charges that individually don't trigger bank alerts, or fraud on a card you rarely check.

Subscription Price Increases

Your streaming service quietly raises its price from $14.99 to $17.99. Your cloud storage goes from $9.99 to $12.99. Each increase is small enough that you might not notice on your statement. But the TimesFM model has a tight expected range for these recurring charges. When the actual amount exceeds the p90 bound, it gets flagged.

Over a year, unnoticed subscription price increases can add up to $200-$400 in incremental spending. Catching each one as it happens lets you decide whether the service is still worth the new price.

Spending Habit Changes

Sometimes the anomaly isn't a single charge — it's a shift in pattern. You started ordering lunch delivery 3 times a week instead of once. Your weekend spending crept from $200 to $350. These gradual changes eventually push daily totals outside the model's expected bands, surfacing them in Clarity's spending insights.

This is different from just looking at monthly totals going up. The model tells you when the change is happening (which days of the week) and how it compares to your historical pattern, giving you specific context for understanding the shift.

Billing Errors

Double charges, incorrect amounts, charges for canceled services — these are more common than most people realize. A 2023 Consumer Reports study found that 49% of Americans had been charged for a subscription they thought was canceled. When these charges appear in your spending data, they push the actual total above the expected range, triggering an anomaly flag.

How It Shows Up in Clarity

Anomaly detection integrates into several existing features:

• Spending insights. The insight bar on the dashboard surfaces anomaly observations: "Your spending on Tuesday was $180 — that's above your typical range of $25-$65 for Tuesdays." Tapping the insight shows the transactions from that day.
• Recurring bills. When a recurring charge comes in above or below the predicted range, it's highlighted on the recurring bills calendar. The predicted amount and actual amount are shown side by side.
• Transaction detail. Individual transactions that contributed to an anomalous day can be flagged in the transaction list, helping you quickly identify which specific charge was unusual.
• AI assistant. Ask "was there anything unusual in my spending this week?" and the assistant references the anomaly detection results, pointing out specific days and amounts that fell outside expected ranges.

The Math Behind the Flags

The detection logic is straightforward once you have the quantile predictions:

• For each day, TimesFM produces a one-step-ahead prediction with p10, p50, and p90 quantiles based on all prior days in the series.
• The actual daily spending total is compared against the p10-p90 range.
• If actual spending exceeds p90, it's flagged as "unusually high." If it falls below p10, it's flagged as "unusually low."
• The severity of the anomaly is proportional to how far outside the band it falls. Spending at 1.5x the p90 value is more notable than spending at 1.05x the p90 value.

By definition, about 20% of days will fall outside the p10-p90 band even under normal conditions. Clarity accounts for this by looking at the magnitude of the deviation, not just whether it crossed the threshold. Minor exceedances are noted quietly; significant ones are surfaced prominently.

Privacy Considerations

Anomaly detection runs on the same aggregated daily totals used for spending forecasts. The TimesFM model sees a sequence of numbers, not individual transactions or merchant names. The comparison between predicted and actual values happens within Clarity — no transaction-level data is sent to the model.

When an anomaly is detected, Clarity then looks at the individual transactions from that day to provide context in the insight. But that lookup happens against your local data, not through the forecasting model.

What It Won't Catch

TimesFM-based anomaly detection has limitations:

• Same-day fraud. If a fraudulent charge happens on a day when you were already spending heavily, the total might not exceed the expected range. Transaction-level fraud detection (which your bank does) is better at catching individual suspicious transactions.
• Very small charges. A $2.99 charge for an app you didn't authorize won't move the needle on daily totals. Small recurring fraud ("cramming") requires transaction-level monitoring, which is a different feature.
• New spending categories. If you've never traveled internationally and then spend $3,000 on a vacation, that's an anomaly by definition — but it's an expected one. The model can't distinguish planned lifestyle changes from unexpected ones.

Anomaly detection is one layer in a broader picture. It works alongside Clarity's spending forecasts, budget tracking, and recurring bill predictions to give you a complete view of whether your finances are on track or veering off course. The forecast tells you where you're headed. The anomaly detection tells you when something unexpected happens along the way.