SEER: Self-Evolving Event Reasoning and Retrieval for Time Series Forecasting

1. Introduction

• Historical numerical observations alone are insufficient in non-stationary regimes where geopolitical, policy, earnings, and supply shocks shape future trajectories.Forecasting supports decisions in finance, weather, energy management, and supply-chain logistics.Source: §1 · pp. 1–2
• Standard retrieval-augmented forecasting can suffer from noisy or misleading news, lack causal attribution, and fail to use forecast errors to adjust context.These shortcomings motivate a closed-loop approach rather than one-shot event retrieval.Source: §1 · p. 2
• SEER converts prediction errors into decoupled retrieval-memory refinement and persistent causal knowledge, with updates indexed only through the current time step.The retrieval memory evaluates whether past queries captured drivers of observed movements; the knowledge base abstracts transferable causal mechanisms.Source: §1 · p. 2
• Across six diverse and volatile time-series benchmarks, SEER consistently outperforms state-of-the-art numerical foundation models, deep sequence architectures, and language-model baselines.The evaluated tasks include component prices, electricity, weather, stocks, and prediction markets; the paper also reports interpretable predictive behaviors.Source: §1 · p. 2

Related Work

• Deep numerical forecasters and time-series foundation models learn temporal patterns or provide zero-shot forecasts, but rely exclusively on numerical inputs.Examples include Autoformer, DLinear, PatchTST, TimesFM 3.0, and Chronos-2.Source: §2 · pp. 2–3
• LLM forecasters integrate language models with static textual information, while missing exogenous event signals and actionable predictive insights can limit their effectiveness.The cited approaches include PromptCast, One-Fits-All, and Time-LLM.Source: §2 · p. 3
• Self-evolving agent research uses reflection, memory selection, and consolidation to distill transferable strategies from past successes and failures.Related work moves beyond archiving raw traces toward structured actionable knowledge.Source: §2 · p. 3
• Unlike adaptive environment-generation frameworks, SEER dynamically enriches the forecasting environment when existing information is insufficient for accurate predictions.The related work describes rigid static testing environments as increasingly bottlenecked by benchmark overfitting.Source: §2 · p. 3

3.1. Problem Setup

• At each cutoff, the task predicts an entity’s outcome over horizon H from its description and historical observations in a look-back window, using only training trajectories whose outcomes are observed by the training boundary.The sequence is split chronologically into training and testing sets to prevent look-ahead bias; the language-model backbone remains frozen, so SEER optimizes its context instead of model parameters.Source: §3.1 · pp. 3–4

3.2. Forecasting with an Evolving Textual Context

• The evolving context combines external events with domain causal knowledge, which are dynamically updatable natural-language components.SEER formulates forecasting as reasoning over structured textual context because endogenous numerical queries may miss exogenous drivers.Source: §3.2 · p. 4
• The context operator expands the event pool through retrieval memory, filters noise through selection memory, and supplies refined events alongside causal knowledge to the forecaster.These components jointly construct the context used for the final forecast.Source: §3.2 · pp. 4–5

3.3. Self-Reflective Evolution

• After outcomes become observable, SEER diagnoses forecast errors and sends them through decoupled updates for retrieval memory and persistent causal knowledge.The separation is intended to prevent interference between inadequate evidence and flawed reasoning; a condensation operator merges similar entries and compresses redundancy as memory approaches capacity.Source: §3.3 · p. 5
• During testing, memory modules are frozen and retrieved events must have occurred by the forecasting cutoff, preventing look-ahead bias.Inference uses the forecasting equation with an initial event pool augmented by retrieval-memory-guided searches.Source: §3.3 · p. 6

4.1. Experimental Setup

• Six domains use chronological learning and evaluation windows with task-specific horizons and metrics; temporal filtering and sequential shared-context updates govern evaluation.The setup uses searches augmented by SEER’s learned retrieval memory and filters retrieved events to prevent future information leakage.Source: §4.1 · pp. 6–7, Table 1 · p. 7

4.2. Ablation Settings

• The ablations progress from historical data alone, to initial-event retrieval, to memory-guided event filtering, and finally to the complete framework with causal knowledge.These configurations isolate the contributions of event retrieval, reflective memory, and causal knowledge.Source: §4.2 · pp. 7–8

4.3. SEER’s Performance on Time-Series Tasks

• 50.2% MSE improvement over the best baseline on Memory 6m and 16.4% error reduction on Electricity; Table 2 compares SEER with time-series baselines.At longer horizons, SEER reduces MSE by 24.3% on Memory and 19.6% on SSD versus the best traditional deep learning baselines.Source: §4.3 · p. 8, Table 2 · p. 8

4.4. Ablation Study And Reasoning Evaluation

• SEER improves over Event RAG by +10.5% for Claude 4.8 Opus and +11.4% for Gemini 3.1 Flash; adding causal knowledge further raises accuracy across backbones.The reported averages compare performance across seven tasks; the complete architecture improves accuracy across all four evaluated backbones.Source: §4.4 · pp. 8–9, Table 4 · p. 10
• Across three pairwise comparisons, Grok-4.20-Reasoning gives SEER the highest win rates against Event RAG and SEER w/o Knowledge on four reasoning dimensions.The judge compares event relevance and plausibility, domain relevance, logic-to-number consistency, and analytical depth while withholding actual numerical outcomes.Source: §4.4 · p. 9, Table 3 · p. 9

4.5. Generalizability to Non-Time-Series Forecasting

• SEER outperforms baseline LLMs on Polymarket event forecasting by Brier Score and Selective Stock Forecasting by Accuracy, extending evaluation beyond numerical time series.These tasks test event resolution and selecting up to six predictable equities from a daily pool of 50.Source: Figure 3 · p. 9, §4.5 · p. 10

4.6. Explainable Forecasting via Causal Reasoning

• At the June 2025 Memory shift, SEER uses distilled causal knowledge to predict a price surge; on selected stocks, its accuracy is 51.2% versus 32.7% for baseline.For mutually selected stocks, SEER and the baseline have similar accuracy (51.7% vs. 48.3%); SEER’s overall gain comes from selecting different, more predictable tickers.Source: §4.6 · pp. 10–11, Figure 4 · p. 11

5. Conclusion

• SEER analyzes correct and incorrect predictions to optimize reflective event retrieval and a persistent domain causal knowledge base.The retrieval memory supports supplementary event retrieval and noise filtering; the knowledge base assists future forecasting.Source: §5 · p. 11
• Across extensive experiments, SEER consistently outperforms state-of-the-art time-series forecasting methods.Source: §5 · p. 11
• SEER generalizes to highly volatile, event-driven tasks, including stock-market and real-world event forecasting.Source: §5 · p. 11

A. Forecast Targets

• Memory targets include server modules across six capacities and two generations, alongside retail kits with specified capacities and data rates.The server modules cover DDR4-2400, DDR4-3200, and DDR5 rates from 4800 to 7200 MT/s; retail kits cover listed DDR4 and DDR5 capacities and rates.Source: A. Forecast Targets · p. 16
• The SSD targets comprise ten flash modules ranging from 32 GB to 8 TB, with MLC and TLC types identified for some modules.Two 4 TB and two 8 TB modules have NAND types not specified in their part names.Source: A. Forecast Targets · p. 16
• Weather targets cover ten US cities, while electricity targets daily maximum and minimum grid loads in Boston, New York, and San Francisco.Source: A. Forecast Targets · p. 16
• Event forecasting uses 100 training and 46 testing events across Tech, Economy, and Politics, with questions, candidate answers, and outcomes retrieved via the Polymarket API.The questions are a subset of ForecastBench; examples include technology milestones, macroeconomic events, and political developments.Source: A. Forecast Targets · p. 16
• The stock target covers 50 large US companies and classifies each stock and cutoff as Up, Neutral, or Down.The listed companies span technology, semiconductors, finance, consumer, healthcare, industry, and energy sectors.Source: A. Forecast Targets · p. 16

A.1. LLM-as-a-Judge Reasoning Evaluation

• Grok-4.20-Reasoning compares candidate models using ground-truth events and each model’s reasoning and predictions, while withholding actual numerical ground truth.The judge makes pairwise comparisons rather than applying standard accuracy metrics.Source: §A.1 · p. 17
• The judge assesses domain relevance, event relevance and plausibility, logic-to-number consistency, and analytical depth.These dimensions cover terminology, causal links to events, consistency between narrative and numerical outputs, and understanding of time-series dynamics.Source: §A.1 · p. 17
• Table 3 reports that SEER consistently outperforms SEER w/o Knowledge and static LLM (Event RAG) in generating logically coherent forecasting rationales.Source: §A.1 · p. 17

B. Prompts and Case Studies

• The learning templates reflect on completed forecasts to update selection and retrieval memories, then use an independent call with the same evidence to distill causal knowledge.Source: B. Prompts and Case Studies · p. 17
• Memory condensation prompts activate when a module exceeds 80% of its character limit, pruning and consolidating its contents.The respective limits are 12,000, 14,000, and 22,000 characters.Source: B. Prompts and Case Studies · p. 17
• The search templates include an initial search, an expanded search guided by retrieval memory to target reflection-identified blind spots, and filtering of prior retrieved events through LEAF fact-checking.Stock searches run daily, while other domains use a seven-day window.Source: B. Prompts and Case Studies · pp. 17–18
• The appendix presents learned causal rules inserted into forecasting prompts: six for memory pricing, 70 for cross-stock forecasting, and seven for city-level hourly electricity loads.Fourteen of the 70 stock rules are shown, retaining their original numbering.Source: B. Prompts and Case Studies · p. 18
• For the 64 GB DDR5 case, SEER applies a learned rule to project an upward squeeze, then uses ×1.780 and ×1.43 multipliers for successive quarter boundaries.The finalized May price serves as the forecast baseline, and the output contains nine forecasted prices.Source: B. Prompts and Case Studies · p. 18
• In the 128 GB DDR5 case, SEER uses events [132], [156], [173], and [66–68] to construct a two-step trajectory from the finalized June price of 875.18.Source: B. Prompts and Case Studies · p. 18
• In the selective-stock case, SEER selects MU Down and AMD Down, both of which closed down next session, while all four same-day baseline predictions were incorrect.The example also shows SEER rejecting candidates based on learned rules, including MRK after an M&A-driven spike and CVX after a single-headline reversal.Source: B. Prompts and Case Studies · p. 18

C. Other Results

• The intervals use 20,000 paired cluster-bootstrap resamples, with forecast targets as resampling units and shared draws across methods.The target unit varies by task: products for Memory and SSD, cutoff days for Stock, and forecast windows for Weather and Electricity.Source: C. Other Results · p. 18
• SEER’s estimated error ranges generally shift lower than baselines, and its 95% confidence intervals support a consistent performance advantage.Table 5 reports intervals for the main forecasting results.Source: C. Other Results · p. 18, Table 5 · p. 19
• 95% confidence intervals for average relative improvements remain positive across evaluated LLM backbones, supporting the robustness of the dual-memory mechanisms.Table 6 reports intervals for component ablations.Source: C. Other Results · p. 18, Table 6 · p. 19
• The learned memory-price rules classify squeeze versus glut using supply, buyer response, supplier guidance, and whether the SKU is trailing- or leading-edge.The rules prioritize correct regime sign, base selection, step timing, and multiplier magnitude; target math is Correct_Base × Precedent_Multiplier.Source: C. Other Results · p. 20
• The 64GB DDR5 forecasting prompt combines historical prices, retrieved events, and causal knowledge distilled from past predictions.Figure 5’s prompt targets nine consecutive months after the 2025-05 cutoff; some events are omitted, retaining only those referenced in the reasoning.Source: C. Other Results · p. 20, Figure 5 · p. 21
• The memory-learning template updates both selection memory and retrieval memory, while a separate knowledge-learning template updates causal forecasting knowledge.The templates adapt placeholders to domain-specific cues and descriptions.Source: Figure 7 · p. 23, Figure 8 · p. 24
• SEER updates causal forecasting knowledge by comparing predicted prices and reasoning with realized trajectories to distill durable rules across parts and periods.The learning prompt asks the model to identify events and mechanisms behind correct or incorrect movements and adjust for overlooked structural dynamics.Source: C. Other Results · p. 24

SEER Memory Condense — [SELECTION MEMORY]

• The selection-memory prompt asks the model to prune flawed or outdated rules, merge overlaps, and retain actionable sector-wide guidance without oversimplifying useful nuance.The updated list is returned in sequentially numbered tags.Source: SEER Memory Condense — [SELECTION… · p. 25

... <SELECTION_MEMORYN> ... </SELECTION_MEMORYN>

• The user prompt supplies the current memory list through the {memory_str} placeholder.Source: ... <SELECTION_MEMORYN> ...… · p. 25

SEER Memory Condense — [RETRIEVAL MEMORY]

• The retrieval-memory prompt generalizes missing signals to sector level, prunes or refines flawed suggestions, merges redundancies, and preserves actionable detail.The resulting list is returned in sequentially numbered tags.Source: SEER Memory Condense — [RETRIEVAL… · p. 25

Event search

• Expanded searches incorporate learned retrieval-memory clues to find missing events without duplicating existing entries; search windows and verified-event filtering are task-specific.Initial stock searches use a daily date window, while other tasks can use a seven-day window; only events verified through multi-agent fact-checking are supplied to forecasting.Source: Figure 10 · p. 27, <RETRIEVAL_MEMORY1> ...… · p. 28, Figure 11 · p. 28

Memory-price forecasting

• Memory-price prompts request exact monthly unit-price forecasts for nine months and direct the model to consult historical forecasting experience.The accompanying knowledge classifies the market as squeeze or glut, then prioritizes regime sign, base selection, and step timing.Source: <RETRIEVAL_MEMORY1> ...… · p. 32

Selective stock forecasting

• Selective stock prompts allow zero to six picks and favor clear structural catalysts aligned with durable learned rules over noisy or mixed signals.Rules include reducing continuation conviction when a catalyst is already priced into a cutoff-session close; in the illustrated case, SEER’s two Down picks were correct while four baseline picks failed.Source: <RETRIEVAL_MEMORY1> ...… · pp. 35–36, Figure 18 · p. 37

Reasoning evaluation

• The LLM judge compares forecast reasoning without access to ground-truth numerical values, assessing domain relevance, event plausibility, and consistency between reasoning and predictions.It is instructed to penalize hallucinated or illogical event links and check whether the described direction and magnitude match the model’s own numbers.Source: <RETRIEVAL_MEMORY1> ...… · p. 38

• The required JSON reports a winner for each criterion and overall preference, plus a 2–4 sentence justification referencing specific criteria.Each winner field accepts Model A, Model B, or Tie.Source: <RETRIEVAL_MEMORY1> ...… · p. 38, Figure 19 · p. 39
• The judge compares two candidate models on domain relevance, event relevance and plausibility, logic-to-number consistency, and analytical depth.Analytical depth concerns understanding time-series dynamics such as trend, volatility, and momentum rather than superficial observations.Source: <RETRIEVAL_MEMORY1> ...… · p. 38, Figure 19 · p. 39
• Ground-truth numerical values are withheld so the pairwise evaluation focuses on reasoning quality rather than outcome bias.The user template injects ground-truth events and outputs from the two candidate models.Source: Figure 19 · p. 39