AI Forecasting in Google BigQuery: Our New Lazy Genius Way

Are you tired answering, how's the traffic going to look like in the next 30 days?
We've got that questions too often, well the common case is marketers often load historical data and do calculation on excel/ sheet.
But what if that process can work with the help of GBQ with ML? The AI.FORECAST function was launched early in 2025 using BigQuery ML's built-in TimesFM model and we finally had chance to try with our datasets.

If you are naturally marketers (just like us) or work with limited data team, it's a truly game changer. AI.FORECAST. It’s a BigQuery function powered by TimesFM model, this lets us do forecasting without having to create, train and manage the model. Just plug in the table, define the time column and the value we want to forecast, and it generates the prediction for us. With confidence intervals that we can define on the SQL query (by default it is set at 95%.

⚡ Wait — Zero Training?
Exactly. Unlike building our own model for example: ARIMA where we have to manually train, evaluate, and maintain models, AI.FORECAST is pre-trained. We’re using Google’s model straight out of the box — zero setup, zero headaches.

This is a foundation model, like GPT but for time series. You don’t tweak it — you just feed it your historical data and it does the rest.

🤹 But It’s “Only” Univariate
Before you get too excited: this only works on one metric at a time.
It’s univariate — which means it forecasts the future of a single column based purely on its past behavior. If you work with a business that drastically ramping up the budget, this might not be an ideal solution for you.

Well, you’ll need to build your own multivariate model in BigQuery ML using ARIMA_PLUS, that’s doable — but it’s not a one-liner (maybe also a right time to change your job title to somewhat data science 👀👀)

Still, for quick, directional forecasts, AI.FORECAST is a massive leap forward.

🧠 The Excel Debate

Dimas and I had a heated debate:

Isn’t this just the same as doing a linear regression in Excel or Sheets?

In short: Nope. Not even close.

Here’s why:

A linear regression is… static.
AI.FORECAST uses transformers — like comparing a sundial to a smartwatch.

🧪 Example: Forecasting Traffic in the next 90 days

TABLE

SELECT
  *
FROM
  AI.FORECAST(
    (
      SELECT
        session_date AS timestamp_col,
        COUNT(session_id) AS sessions_count_data_col 
      FROM
        `your table`
      
      GROUP BY
        session_date
      ORDER BY
        session_date
    ),
    data_col => 'sessions_count_data_col',
    timestamp_col => 'timestamp_col',
    horizon => 90
  );

That’s it.

BigQuery handles everything:

✅ Modeling

✅Validation

✅ Seasonality detection

✅Scaling across thousands of time series

📦 Output You Get

TABLE

forecast_timestamp Future date
forecast_value Predicted value (mean of prediction interval)
prediction_interval_lower_bound Lower bound of the forecast
prediction_interval_upper_bound Upper bound of the forecast
confidence_level Confidence level used (by default it is 95%)
ai_forecast_status Error info (blank = all good)

📈 Traffic Forecast – AI vs Actual

🔍 Insight: Despite being a univariate model, AI.FORECAST holds strong predictive power, staying within ~10% of actual values.

🔐 One Last Thing — Security

You know what’s even better than pasting it to AI tools?

Keeping it inside your secure data stack.
Unlike ChatGPT or other AI tools, by using this, it keeps your data in place. No exporting to third-party models.

🧯 Limitations to Know

Before you go full AI-hype, here are the current boundaries:

• ❗ Only support univariate
• ⌛ Max 512 time points
• ✅ Needs minimum 3 data points
• 🔢 Only works with numeric types (INT64, FLOAT64, etc.)
• 🧠 Powered only by TimesFM

🧩 When (and When Not) to Use It

✅ Use AI.FORECAST when you want to:

• Predict traffic, bookings, revenue, or occupancy
• Forecast per region/city/property
• Avoid building and maintaining ML pipelines
• Impress your client with real insights, not hand-waving

🚫 Don’t use it when:

• You need multiple variables to influence the forecast
• Your data is chaotic or sparsely collected
• You need business logic baked into your forecast