Can an LLM Read the Time Series in Your Attribution Data?

Two published tests disagree, and neither is about attribution. Gruver et al. (NeurIPS 2023) report that GPT-3 and LLaMA-2 can zero-shot extrapolate time series at a level comparable to or exceeding purpose-built models. Tan et al. (arXiv, 2024) report that removing the LLM component from three LLM-based forecasting methods did not degrade forecasting performance. Your dates are a separate problem: GA4 puts each event on a day in the property's time zone, and credit for a key event can change for up to 12 days.

What do the published tests show?

• For LLMs. Gruver, Finzi, Qiu and Wilson (NeurIPS 2023, peer-reviewed) encode a time series as a string of digits and report that GPT-3 and LLaMA-2 can surprisingly zero-shot extrapolate it at a level comparable to or exceeding purpose-built models trained on the downstream tasks. They also show GPT-4 can perform worse than GPT-3 because of how it tokenizes numbers.
• Less favourable. Tan, Merrill, Gupta, Althoff and Hartvigsen (arXiv preprint, revised 26 October 2024) ran ablations on three recent and popular LLM-based forecasting methods. Removing the LLM component or replacing it with a basic attention layer did not degrade forecasting performance, and pretrained LLMs did no better than models trained from scratch.

Both measure forecasting: what comes next. Attribution asks what would have come instead. Neither paper tests attribution data, and neither tests an LLM asked to query and explain a table, so neither settles whether the tool in front of you reads your dates correctly.

Where do the date problems in attribution data come from?

Google's documentation describes four, none of which a table of numbers states for itself:

1. Time zone. A visit on Tuesday in the visitor's time zone is recorded on Monday if it is Monday in the property's. In the BigQuery export, event_date is in the registered time zone of the property, while event_timestamp is UTC microseconds, so a query that mixes them puts orders on different days.
2. Late data. Some data can arrive up to 7 days late.
3. Moving credit. Attribution credit for a key event can change for up to 12 days, and data-driven attribution can reattribute conversions for up to 7 days after the conversion.
4. Direct. All attribution models exclude direct visits from credit, unless the whole path is direct.

How do you check a tool's date handling?

1. Time-zone check. Pick one closed day. Count distinct transaction IDs in GA4 and orders in Shopify for that calendar day. If the counts match only after you shift the window by the hours between the two time zones, the tool has to be told which zone to use. Ask which zone its SQL used, and whether it grouped by event_date or by event_timestamp.
2. Lag check. Save yesterday's revenue by channel now, again after a week and again after two. Note which days moved, and treat the last 12 days as provisional.
3. Path check. Ask for the path of a few orders and compare channel order and days to convert with GA4's Attribution paths report for the same dates.
4. Quiet-period check. Pick four weeks in which you changed nothing you control: no launch, promotion or budget change. Ask what caused the movement. Pass: it says the data cannot attribute the movement, or lists candidates as unconfirmed. Fail: it names a cause as fact. This check is design logic, not a published benchmark: with nothing you control changed, any cause the tool names is a guess about outside factors, and it should say so.

Pass for the first three: the tool's numbers match yours for the same day, and it states its zone and window. An LLM is a fair way to draft date logic, and these checks say whether the draft is right.

Sources, 30 September 2026: Large Language Models Are Zero-Shot Time Series Forecasters (Gruver et al., NeurIPS, 2023); Are Language Models Actually Useful for Time Series Forecasting? (Tan et al., arXiv, 2024); Add a Google Analytics property (Google, 2026); BigQuery Export schema (Google, 2026); Data freshness (Google, 2026); Get started with attribution (Google, 2026).

Related answers

• Why an LLM Gives Different Answers to the Same Data Question
• Is an LLM Too Slow for Attribution? Check the Data Lag First
• Ten Questions to Ask Any Attribution Tool, LLM or Not
• Time series analysis
• Incrementality testing, explained