First Reasoning, Then Forecasting

• First Reasoning, Then Forecasting is a principle where prediction follows a structured process of historical data reconstruction and uncertainty modeling.
• It distinguishes time-aware reasoning from look-ahead bias by ensuring forecasts are based solely on evidence available prior to the prediction point.
• Emerging agentic architectures integrate iterative reasoning, tool-based interventions, and explicit memory to improve forecast calibration and robustness.

“First reasoning, then forecasting” denotes a forecasting principle in which prediction is treated as the final stage of a prior inferential process: the forecaster must first reconstruct the historically available information set, identify relevant structure, weigh competing evidence, and reason under uncertainty, and only then emit a forecast. In contemporary work, the phrase is used both narrowly, to distinguish historically grounded prediction from hindsight-contaminated output in LLMs, and broadly, to describe agentic and benchmarked forecasting systems in which retrieval, decomposition, calibration, evidence selection, and explicit uncertainty handling precede prediction (Merchant et al., 25 Jun 2026, Hosni et al., 2017).

1. Definitional core and scope

In its most specific recent formulation, forecasting is the problem of mapping a historical information set to a future outcome, but an LLM can answer such a prompt in two different ways. The first is time-aware reasoning, in which the model reasons as of the historical date and produces the option that would have been most plausible then. The second is look-ahead-biased reasoning, in which the model implicitly uses knowledge of what actually happened later. The distinction matters because an answer can coincide with the realized future while still failing as a historically grounded forecast (Merchant et al., 25 Jun 2026).

This definition generalizes beyond one task family. In temporal knowledge graph forecasting, the query is explicitly constrained to histories before the query time, such as $$\mathcal{G}_{<t^q}=\{\mathcal{G}_1,\mathcal{G}_2,\cdots,\mathcal{G}_{t^q-1}\}$$, and the system is expected to predict the missing future entity from that restricted evidence rather than from unrestricted world knowledge (Xia et al., 2024). In benchmark design, the same idea appears as the requirement that a model receive a question $Q$ and return both an answer and a confidence score, $M(Q)\to(A,C)$, where the answer is defined relative to knowledge available up to a cutoff time $t_{train}$ rather than post-resolution information (Yuan et al., 27 Feb 2025).

A common implication across these formulations is that a forecast is not identified with the final output token, number, or probability. It is the endpoint of a reasoning process whose validity depends on temporal admissibility, evidential grounding, and uncertainty representation. This suggests that “forecast accuracy” and “forecast validity” are separable notions: a prediction may be numerically correct for the wrong historical reasons, or numerically imperfect while still reflecting better historical reasoning (Merchant et al., 25 Jun 2026, Yuan et al., 27 Feb 2025).

2. Historical and methodological antecedents

The reasoning-first view has clear methodological antecedents in the philosophy of forecasting. One influential formulation contrasts a reductionist, first-principles approach with a naive inductivist or big-data approach and argues that neither is sufficient on its own. In weather forecasting, Richardson’s first-principles vision was broadly correct in spirit, but von Neumann and Charney showed that forecasting needed effective equations at the right level of description rather than maximal microscopic detail. The central lesson is that more detail does not automatically improve prediction; forecasting often depends on suppressing irrelevant fast variables and selecting the appropriate abstraction (Hosni et al., 2017).

The same paper formalizes the limits of naive analogy-based prediction. If one seeks a past state $\mathbf{x}_k$ close to the present state $\mathbf{x}_M$ such that $|\mathbf{x}_k-\mathbf{x}_M|<\epsilon$ and then predicts $\mathbf{x}_{M+n}\approx \mathbf{x}_{k+n}$, the required data scale exponentially with effective dimension. The expected recurrence time scales like $O(\epsilon^{-D})$, which makes “data alone” forecasting rapidly impractical in high-dimensional systems (Hosni et al., 2017). The point is methodological rather than purely mathematical: forecasting requires conceptual organization of information, not only larger historical archives.

Logic-based meteorological work reaches a similar conclusion from a different direction. MeteoLOG treats forecasting as reasoning over time-stamped, model-derived, and potentially conflicting assertions, represented as labeled assertional maps $\langle \lambda,\tau^t\rangle:Q(x,y,z,\tau^r,q)$, where model identity, generation time, and source accuracy determine priority. The Tournament algorithm converts such ordered metarules into a defeasible theory, discards outdated maps, compares conflicting assertions, and outputs a theory on which an automatic reasoner can operate before any user-facing bulletin is produced (Cristani et al., 2019). In this formulation, the forecast bulletin is explicitly the surface form of a deeper conflict-resolution and priority-reasoning pipeline.

3. Reasoning within forecasting models

Recent LLM work has moved the reasoning-first principle inside the model itself. In “Forecasting With LLMs: Improved Generalization Through Feature Steering,” sparse autoencoders are used to decompose hidden activations into sparse features associated with temporal concepts. Candidate features correlated with time-aware reasoning and with look-ahead-biased reasoning were identified on prediction-market prompts, and causal intervention was then performed by amplifying a selected latent feature via

$Q$0

Across M&A and pharmaceutical growth-driver tasks, amplifying time-awareness features substantially reduced look-ahead-biased outputs while preserving general reasoning performance on MMLU CoT and MMLU-Pro CoT; steering the candidate look-ahead-bias features did not reliably reduce bias (Merchant et al., 25 Jun 2026). The paper’s empirical emphasis is therefore on causal usefulness of time-aware features, not on the claim that every discovered feature is itself causal.

At the prompt level, however, explicit reasoning is not uniformly helpful. A separate study of LLM time-series forecasting found that no prompting method was universally best and that a context-only baseline could match or exceed reasoning-heavy prompts such as zero-shot CoT, one-shot CoT, PaS+, SARIMA-style decomposition, and LSTPrompt. The reported failure modes were procedural noncompliance, arithmetic error, semantic misunderstanding, and incomplete outputs, including cases in which LST prompting generated only five of the required six predictions or SARIMA-style decomposition inflated component estimates by failing to subtract trend before estimating seasonality or residuals (Yang, 8 Feb 2025). This complicates any simple equation of “more visible reasoning” with “better forecasting.”

Benchmarking work refines that picture. ReC4TS, a benchmark for zero-shot time-series forecasting, reports that self-consistency is the most effective test-time reasoning strategy, that Group Relative Policy Optimization is more suitable than other post-training approaches for incentivizing reasoning ability, and that multimodal forecasting benefits more from reasoning than unimodal forecasting. The benchmark further reports that self-consistency outperforms direct System 1 in roughly 60%–80% of cases and that, in long-term forecasting, CoT, self-consistency, and self-correction beat the direct baseline at rates of 52.08%, 58.33%, and 54.17%, respectively (Liu et al., 27 Feb 2025). By contrast, “Context information can be more important than reasoning for time series forecasting with a LLM” argues that proper context can be more crucial than asking for a specific type of reasoning (Yang, 8 Feb 2025). Taken together, these results suggest that the benefit of reasoning is architecture-, task-, and protocol-dependent.

A further line of work asks whether successful forecasting may already reflect implicit reasoning rather than explicit chain-of-thought. Controlled out-of-distribution experiments on composition, comparison, and inverse search show that certain linear, MLP-based, and patch-based Transformer models generalize beyond training distributions in ways consistent with underexplored reasoning capabilities. On synthetic tasks, DLinear was best on addition and subtraction composition, NHITS on multiplication and inverse search among trained models, and TimesFM and Chronos were especially strong in zero-shot comparison and inverse search (Potosnak et al., 2024). A different response is to make reasoning explicit and train it directly: Time-R1 uses supervised warmup plus reinforcement learning, explicit > ... <answer> ... </answer> formatting, and a multi-objective reward over format, length, numerical accuracy, seasonal-trend decomposition, and structural similarity, with GRIP as the policy-optimization mechanism (Luo et al., 12 Jun 2025).

4. Agentic and tool-mediated forecasting architectures

The reasoning-first idea becomes most explicit in agentic architectures. A position paper on Agentic Time Series Forecasting (ATSF) reframes forecasting as an iterative process built from perception, planning, action, reflection, and memory, organized as

$Q$1

Its claim is that real forecasting is rarely a static, single-pass mapping from historical observations to future values; instead, it requires relevant-information selection, task decomposition, tool use, revision, and experience accumulation (Cheng et al., 2 Feb 2026).

Several concrete systems instantiate this view. Reasoning and Tools for Forecasting (RTF) uses hierarchical ReAct agents with Google Search API and a Python interpreter. On 201 binary Manifold Markets questions scraped on April 15, 2024 and resolving within the following two weeks, the search tool was restricted to information before the cutoff date to avoid leakage. The resulting agent matched or slightly exceeded crowd performance: RTF Mean of 3 achieved 73.9% accuracy versus the crowd’s 73.8%, while RTF Median of 3 achieved Brier 0.169 versus the crowd’s 0.172 (Hsieh et al., 2024). The claim is not that a base LM “knows” the future, but that retrieval and computation can ground a forecasting workflow.

Other systems specialize the same sequence. Chain-of-History (CoH) for temporal knowledge graph forecasting forces the model to select important first-order histories, expand them into higher-order chains, and forecast only after this iterative select-and-expand process; on ICEWS14, ICEWS18, and ICEWS05-15, it improved over a prior ICL-based LLM method and also improved RE-NET, RE-GCN, and TiRGN when used as a plug-and-play module (Xia et al., 2024). TimeFore, introduced with ODTQA-FoRe, separates future-oriented tabular QA into retrieval, forecasting, and answer synthesis. The Retriever summarizes a question into a caption-like form and uses direct match or BM25 over table captions; the Forecaster converts a 24-month historical window into a time series, uses TimesNet for imputation and TimeXer for forecasting, and produces a 12-month forecast for 2024; the Analyzer classifies the query as direct forecasting or forecast-based reasoning and then synthesizes the final answer (Wang et al., 1 Jun 2026).

Recent multimodal and financial systems go further by making intermediate reasoning objects explicit. Nexus decomposes prediction into a Historical Context Agent, Macro-Reasoning Agent, Micro-Reasoning Agent, Forecast Synthesizer Agent, and Calibration Agent, and on post-cutoff Zillow and stock datasets it improved over direct CoT baselines; for Gemini-3.1-Pro on Zillow, average MAPE moved from 0.0423 to 0.0361 and RMSE from 63.1264 to 53.4620 (Das et al., 14 May 2026). StockR1 makes the intermediate object a structured forecast action with fields such as direction, end-point movement, peak timing, volatility level, and drawdown severity; this action conditions a time-series decoder, and reinforcement learning optimizes answer validity, action accuracy, forecast precision, and action-trajectory consistency. On a 10-year financial benchmark, the method reports reasoning-accuracy improvements of 17.7% at 4B and 25.9% at 8B (Chen et al., 21 May 2026). ForecastCompass instead places the reasoning stage in memory: it organizes forecasting experience through a hierarchical taxonomy, with factor memory for reusable predictive dimensions and reasoning memory for calibration principles, and reports improved Brier and ECE on Prophet Arena and FutureX with GPT-5-mini and Gemini-2.5-Flash (Chang et al., 29 May 2026).

5. Evaluation, calibration, and argumentative coherence

Once reasoning is treated as part of forecasting, evaluation must move beyond point accuracy alone. FOReCAst is built around this premise: it evaluates Boolean questions, timeframe prediction, and quantity estimation while making confidence assessment first-class. The benchmark contains 2,256 questions, split 65% train, 10% validation, and 25% test, and computes calibration with a modified Brier score for Boolean tasks and CRPS for timeframe and quantity tasks. Its main empirical conclusion is that forecasting remains hard for current LLMs, especially calibration, and that point prediction quality and confidence quality do not reliably move together (Yuan et al., 27 Feb 2025).

A complementary move is to benchmark the reasoning process itself. Bench to the Future 2 (BTF-2) contains 1,417 pastcasting questions, a frozen corpus of about 16.2 million scraped web documents, and full agent traces including searches, page reads, thoughts, and final rationales. It can detect 0.004 Brier differences and separates research from judgment by comparing end-to-end agents with fixed-evidence judging. On the simple prompt, Opus 4.6 achieved 0.131 Brier, but a stronger constructed forecaster reached 0.119, a 0.011 improvement; the identified strategic-reasoning differences centered on pre-mortem analysis, blind-spot awareness, and wildcard or black-swan consideration (Liptay et al., 28 Apr 2026). The benchmark’s methodological claim is that exposing reasoning traces enables analysis of why a forecaster is stronger, not only whether it is stronger.

TFRBench pushes this farther by explicitly formalizing reasoning as an intermediate variable:

$Q$2

Its multi-agent generation pipeline uses Search, Reasoning, Verifier, Forecasting, and Summary agents, with a generate-verify-refine loop and numerical acceptance criteria based on MASE. Spanning ten datasets across five domains, the benchmark reports that prompting LLMs with generated reasoning traces improves forecasting accuracy from about 40.2% to 56.6%, while off-the-shelf LLMs often fail on both reasoning quality and numerical prediction (Ahamed et al., 7 Apr 2026). The paper’s central distinction is between causally grounded reasoning traces and merely fluent rationales.

Argumentation-based forecasting formalizes the same idea in symbolic terms. Forecasting Argumentation Frameworks (FAFs) organize forecasting as a sequence of update frameworks containing a proposal forecast, increase and decrease amendment arguments, pro and con arguments, votes, and individual forecasts; the final group forecast is a weighted mean using historical Brier-derived weights, $Q$3 (Irwin et al., 2022). “Argumentatively Coherent Judgmental Forecasting” defines argumentative coherence by requiring that the direction of a forecaster’s probability align with the argumentative strength $Q$4 of the reasons they give. Empirically, filtering incoherent predictions improved forecasting accuracy for both humans and LLM-based forecasters, while crowd experiments showed that users do not generally align with this coherence property (Gorur et al., 30 Jul 2025). These frameworks treat reasoning not as optional explanation, but as a constraint on admissible forecasts.

6. Limits, controversies, and emerging research directions

The literature does not support a simple claim that explicit reasoning always improves forecasting. One set of results shows that time-awareness feature steering reduces look-ahead bias while preserving broader reasoning performance, but also that steering candidate look-ahead-bias features did not work reliably, partly because the discovery proxy was noisy and partly because the discovery task used multiple choice while the intervention task used free-form generation (Merchant et al., 25 Jun 2026). Another line reports that context-only prompting can equal or exceed explicit reasoning prompts and catalogs failures such as arithmetic mistakes, procedural noncompliance, and semantic misunderstanding (Yang, 8 Feb 2025). A third shows that unverified reasoning can induce narrative bias, especially in stochastic domains, and may worsen forecasting rather than improve it (Ahamed et al., 7 Apr 2026).

A second controversy concerns what kind of reasoning is actually useful. ReC4TS finds that self-consistency is the most effective test-time strategy and that most System 2 reasoning models do not reliably improve zero-shot TSF, with DeepSeek-R1 as the main exception (Liu et al., 27 Feb 2025). BTF-2 shows that better performance may come from better research strategy rather than better judgment over fixed evidence, since Gemini 3.1 Pro judged fixed evidence better while Opus 4.6 performed better end-to-end when it controlled its own research (Liptay et al., 28 Apr 2026). This suggests that “reasoning” in forecasting is not a single capacity; it includes evidence search, temporal filtering, probabilistic judgment, calibration, and self-critique.

A third open problem is reliability under uncertainty and adaptation over time. Agentic forecasting papers identify memory design, toolkit infrastructure, efficiency, uncertainty-aware control, safety, privacy, and accountability as unresolved issues (Cheng et al., 2 Feb 2026). ForecastCompass notes that interpretability and verification of stored memories remain difficult, even when memory is explicitly factor-centric (Chang et al., 29 May 2026). StockR1 responds by reweighting reinforcement learning with a sample-level uncertainty scalar, thereby down-weighting high-volatility, low-reliability samples during policy updates (Chen et al., 21 May 2026). This suggests a broader convergence: future work is likely to treat calibrated uncertainty, memory revision, and evidence provenance as part of reasoning itself rather than as post-hoc additions to a predictor.

Across these debates, the stable core of “first reasoning, then forecasting” is not a commitment to verbose chain-of-thought. It is the requirement that prediction be historically admissible, causally or evidentially grounded, structurally organized, and uncertainty-aware before it is evaluated as a forecast. In recent forecasting research, that requirement appears in temporal feature steering, chain selection in TKGs, ReAct-style search and simulation, multi-agent decomposition, action-conditioned time-series decoding, factor memory, coherence constraints, and reasoning-aware benchmarks (Merchant et al., 25 Jun 2026, Xia et al., 2024, Hsieh et al., 2024, Ahamed et al., 7 Apr 2026).