How much does workflow architecture matter compared to raw model capability in forecasting?
This explores whether the way you structure the forecasting pipeline — splitting the problem into stages — matters more than how powerful the underlying model is.
This explores whether workflow architecture beats raw model capability in forecasting, and the corpus comes down hard on the side of architecture. The clearest result: LLMs are better forecasters than we give them credit for, but only when the workflow separates numerical reasoning from contextual reasoning. Cram both into one monolithic prompt and the latent ability stays hidden; decompose the task and it surfaces Can LLMs actually forecast time series better than we think?. The Nexus system makes this concrete — it splits forecasting into contextualization, a dual macro/micro outlook, and a synthesis stage, and beats both pure time-series models and pure LLM baselines precisely because no single model is forced to juggle event-driven reasoning and number-crunching at once Can decomposing forecasting into stages unlock numerical and contextual reasoning?.
What's interesting is that this isn't a quirk of forecasting — it's a recurring pattern. Separating the 'planner' from the 'solver' in multi-step reasoning improves accuracy across domains, and the decomposition skill even transfers to new problems while raw solving ability doesn't Does separating planning from execution improve reasoning accuracy?. The lesson is that interference between two cognitive jobs degrades both; giving each its own slot in the workflow removes the bottleneck. That's architecture buying you capability you already had but couldn't access.
The deeper point is that 'capability' often isn't a property of the model at all but of the system around it. Routing queries to specialized models per semantic cluster outperforms a single frontier model — ten small models with a good router beat GPT-4.1 and 4.5, suggesting selection is a stronger lever than scaling Can routing beat building one better model?. The same theme runs through recommender research, where problem-specific design choices and constraints beat deeper, higher-capacity networks What architectural choices actually improve recommender system performance?. Over and over, where you put the structure matters more than how big the model is.
There's a useful caveat hiding here, though: architecture isn't a free lunch. Weak-model committees only match strong models when there's an external signal — a test, a proof, a verifiable check — to select the right answer from the pile; sampling alone amplifies coverage but can't pick the winner When can weak models match strong model performance?. Forecasting's version of that 'soundness signal' is the numerical-contextual separation itself, which is why the decomposition works rather than just adding stages for their own sake.
So the answer to 'how much does it matter' is: a lot, and in a way that should change how you think about the problem. The frontier-model arms race is one axis; the orchestration of weaker or general models into the right stages is a parallel axis that's often cheaper and sometimes wins outright. If you're trying to forecast, the most leveraged move may not be a bigger model but a workflow that stops asking one model to do two incompatible kinds of reasoning at the same time.
Sources 6 notes
LLMs have stronger intrinsic forecasting ability than recognized, but only when workflows separate numerical reasoning from contextual reasoning. Monolithic prompting obscures this capability; structured decomposition surfaces it.
Nexus outperforms pure TSFM and LLM baselines on real-world datasets by decomposing forecasting into contextualization, dual-resolution macro/micro outlook, and synthesis stages. Separating numerical extrapolation from event-driven contextual reasoning avoids forcing one model to handle both simultaneously.
Modular architectures with separate decomposer and solver models outperform monolithic LLMs, with decomposition ability transferring across domains while solving ability does not. The separation prevents planning-execution interference and produces more generalizable skills.
Avengers-Pro achieves 7% higher accuracy than GPT-5-medium by routing queries to optimal models per semantic cluster, or matches its performance at 27% lower cost. Ten 7B models with routing previously surpassed GPT-4.1 and 4.5, suggesting selection is a stronger lever than scaling.
Research shows that architectural choices like removing hidden layers, enforcing constraints on self-similarity, and using appropriate likelihood functions deliver better results than deeper or more complex models. This suggests that problem-specific design decisions matter more than raw representational capacity.
Sampling alone amplifies coverage but cannot select correct solutions. Reliable performance matching requires external soundness signals—tests, proofs, or type checks—that convert latent correct proposals into actual selections.