Ask an AI to predict next week versus next year — does looking further ahead reliably make it worse?

This explores whether predicting further out reliably makes LLM forecasts worse. The collection has no study that tests this directly. None of the retrieved notes compares accuracy at short and long horizons, so there is no measured decay curve to report. What the notes do give is indirect evidence, and it leans away from 'systematic degradation' and toward 'the result depends on how the forecast is built'.

The closest hint is that forecasting quality seems to depend more on structure than on raw model strength. One note finds that LLMs forecast better than expected once the workflow separates numerical reasoning from contextual reasoning, and that a single monolithic prompt hides this ability (Can LLMs actually forecast time series better than we think?). Another builds horizon into the design: its system produces a dual-resolution macro and micro outlook before synthesizing a forecast (Can decomposing forecasting into stages unlock numerical and contextual reasoning?). That treats long-range and short-range views as separate jobs. It suggests horizon is something to engineer around, but the summary doesn't report accuracy at each horizon, so it doesn't show that degradation happens or doesn't.

The evidence on real-world forecasting also blurs any horizon effect. A retrieval-augmented system reached near-parity with competitive human forecasters on questions published after the model's training cutoff, and newer model generations improved without any forecasting-specific tuning (Can retrieval-augmented language models forecast like human experts?). If accuracy shifts with model generation as much as with anything else, a horizon effect would be hard to isolate from it. In founder-success prediction, several LLMs beat human experts, but only because the human bar there was modest (Can language models beat human venture capital experts?). That shows a forward-looking task can be within reach. It says nothing about how the gap changes as the target moves further out.

Two other notes matter for how you would test the question. First, a fixed seed and zero temperature give a repeatable answer, but that answer is still one draw from the model's distribution (Does setting temperature to zero actually make LLM outputs reliable?). A horizon curve built from single runs could mistake noise for a trend. Second, 'long horizon' has a different meaning in agent work. On very long optimization tasks, success was best predicted by persistence in feedback loops, and most models quit early or wasted their budget (What predicts success in ultra-long-horizon agent tasks?). That concerns long tasks, not long predictions, but it shows that failures at long horizons can come from how a model behaves over time rather than from the prediction itself being harder.

The corpus can't say yes or no. A direct answer needs a horizon-by-horizon comparison that controls for model generation and workflow design, and that comparison isn't in this collection.