LLMs and Time-Series Forecasting
An honest look at what LLMs can and can't do with time-series data — what the benchmarks actually test, where the genuine use case is, and when to reach for statistical methods instead.
Every six months there is a paper claiming that large language models can forecast time series competitively with, or better than, statistical baselines. The claim is technically correct: LLMs can produce numeric predictions for future time steps. The benchmark numbers are often impressive enough to generate attention. The question before the next wave of enthusiasm is what is actually happening, and in which specific circumstances it represents progress.
Why skepticism should be the default
Statistical time-series methods, including ARIMA, ETS, Prophet, neural variants like N-BEATS, and temporal fusion transformers, were designed specifically for sequential numeric data. Their inductive biases encode the properties of time series: stationarity, seasonality, autocorrelation structure, trend decomposition. A well-tuned ARIMA model on a stationary series is exploiting specific mathematical properties of that series, properties that LLMs were not trained to recognize or exploit.
LLMs were trained on text. The representation of "next month's sales will be 4,200 units" is fundamentally different from what an ARIMA model does when it models the autoregressive structure of a time series. The LLM has likely seen time series data in its training corpus, including financial tables, spreadsheets embedded in web pages, and academic papers discussing datasets, but it has not been trained to minimize forecast error on sequential numeric data. It has been trained to predict the next token in a text, which sometimes happens to be a number that follows other numbers.
When LLMs perform well on time-series benchmarks, the performance is typically explained by one of three factors: the LLM is exploiting patterns that appear in text about the series (names, categories, seasonal signals that are lexically present), the benchmark is testing a task that is language-forward rather than statistics-forward, or the baseline is weak enough that the LLM's general pattern-matching capability is competitive.
What the benchmarks actually test
A portion of the LLM time-series literature tests models on public datasets such as M4, ETTh1, and Exchange Rate, where the LLM receives a sequence of numeric values and predicts future values. On these benchmarks, recent results have shown LLMs performing comparably to, and in some cases better than, statistical baselines.
The honest read on this: these are benchmarks, not production forecasting tasks. The datasets are clean, the series are stationary or have well-behaved trend and seasonality, and the competitive baselines are not always the strongest available. When the comparison is against a properly tuned temporal fusion transformer or N-BEATS model with appropriate hyperparameter search, the LLM advantage typically shrinks or reverses.
More importantly, these benchmarks don't test the scenarios that make forecasting hard in practice: irregular resampling, missing data, structural breaks due to external events, and the requirement to combine numeric signals with contextual information that is not present in the historical series. They test how well a model can extrapolate from a clean numeric sequence, which is the part of the problem that statistical methods handle well.
There is also a representation issue that benchmark tables often hide. To give a time series to an LLM, you have to turn numeric sequence data into tokens: comma-separated values, serialized tables, natural-language descriptions, or some hybrid format. That conversion is not neutral. It changes precision, context length, scale, and sometimes even the apparent regularity of the sequence. A model that performs well on a serialized benchmark may be learning artifacts of the representation as much as properties of the underlying series.
Where LLMs help in time-series contexts
The useful case is narrower and more interesting than "LLMs can forecast time series." It is: LLMs can reason about time series in ways that statistical models cannot, specifically when the relevant signal is contextual rather than statistical.
Consider zero-shot anomaly detection with context injection. A statistical model identifies anomalies by comparing observations to a learned distribution of normal behavior. A value is anomalous if it falls outside the expected range given the recent history. This works for anomalies that are statistically unusual. It doesn't work for anomalies that are expected given external context but statistically normal given the recent history.
An example: a retail demand series shows a 30% spike in a specific product category. A statistical model may not flag this as anomalous because the series has had previous spikes of similar magnitude. An LLM that is given the same series along with the context "a prominent influencer posted about this product category three days ago, triggering a viral trend" can reason that the spike is causally explained, expected to be short-lived, and likely to be followed by a demand pullback as the trend fades. This is not a statistical inference. It's a causal inference grounded in external knowledge.
This is the complementary capability: statistical models are better at extrapolating within-distribution patterns; LLMs are better at reasoning about why patterns might break and how they might break. Combining them, using statistical models for baseline prediction and LLMs for contextual anomaly explanation and scenario reasoning, is more defensible than either approach alone.
In that architecture, the LLM should usually sit around the forecast rather than inside the core forecasting loop. It can generate hypotheses about regime changes, summarize external events, explain anomalies to operators, and propose scenario adjustments that a human or a separate model can evaluate. That is different from asking it to emit the final numeric forecast directly. The first use gives the LLM the part of the problem where language and context matter; the second asks it to compete with methods whose entire purpose is numeric extrapolation.
Forecasts and explanations are different
A related observation: LLMs are more useful for explaining time-series patterns than for forecasting them. Given a series with an unusual pattern and contextual information about the surrounding environment, an LLM can often produce a coherent narrative explanation of the pattern that a statistical model cannot. "Sales declined in Q3 despite the seasonal pattern suggesting growth, likely because a competitor launched a competing product in mid-June and the category experienced sentiment-driven purchasing delays." This explanation is not a forecast. But it is valuable for decision-making, and it's something a statistical model cannot produce.
The distinction matters because it clarifies what problem you're actually trying to solve. If the goal is a point estimate or a confidence interval for future values, statistical methods are generally more reliable. If the goal is understanding the causal structure of a pattern, why it happened, whether it's likely to continue, what external factors are relevant, LLMs provide something useful.
The practical question
For a new forecasting project, the practical question is whether the hard part is statistical pattern recognition or contextual reasoning. If you're forecasting a clean, high-frequency series with stable structure, such as daily traffic to a web property, weekly product sales in a mature market, or monthly energy consumption in a predictable facility, a statistical model is the right starting point. An LLM is not going to meaningfully improve on a well-tuned ARIMA or Prophet model, and it's going to be significantly more expensive per inference.
If you're forecasting in an environment where external signals are important, including new market entry, product launches, policy changes, macroeconomic shifts, and social dynamics, and if those signals can be represented in natural language and incorporated as context, then an LLM layer can help. Not as a replacement for the statistical baseline, but as a complementary component that adds the contextual reasoning the statistical model doesn't have.
The question I ask before adding an LLM to a forecasting pipeline: is the thing that's making this series hard to forecast a statistical property (complex seasonality, non-stationarity, high noise) or a contextual property (external events, behavioral dynamics, regime changes)? Statistical properties call for better statistical models. Contextual properties call for LLMs. Reaching for LLMs to solve statistical problems is the mistake most commonly made, and it's a mistake that becomes visible in production when the LLM confidently produces plausible-sounding forecasts that are statistically incoherent. The useful boundary is not whether the model can output a number. It is whether the number was the hard part.