midgley Release v0.5.2 Model v1.4 Finlight-LLM

LLM-Augmented Unleaded Gasoline, NOAA Weather & Alternative Physical Data Engine

Overview National Wholesale
Diesel ULSD Fill-Up Advisor Telemetry & Map Math Guide Citations GitHub

Academic Literature & Citation Ledger

Midgley bridges empirical econometric modeling with cutting-edge academic AI research. This ledger tracks all peer-reviewed papers, foundational time-series theorems, and algorithm designs implemented directly within our forecasting architecture.

RESEARCH_CITATIONS.md View Math Guide
12

Implemented Papers

100%

Open-Access / arXiv

6

Core Python Modules

0.1069

Out-of-Time MAE ($/gal)

#01

When Does Context Routing Help? A Systematic Study of Multi-Modal Fusion in Time Series Forecasting

Aug 2026 arXiv:2608.25128v1 PDF

Authors

Ruizhe Zhou, Gaoyuan Du, Xiaoyang Liu, et al.

Target Modules

src/feature_engineering.py

Methodology Type

Pre-Training Diagnostic & RBU Bound

Key Theorem & Mathematical Formulation

$$\mathrm{MMSE}(X_{t+h} \mid X_t, C) = \sigma^2 (1 - \rho_h^2) \cdot 2^{-2\delta}, \quad \mathrm{RBU} = 1 - 2^{-2\delta}$$

Calculates rolling autocorrelation $\rho_h = \mathrm{Corr}(X_t, X_{t+h})$. When $\rho_h > 0.95$ (calm baseline markets), last-value shortcuts dominate and LLM qualitative context fusion is suppressed to prevent hallucination. When $\rho_h \le 0.95$ (exogenous shock events), event memory decay and shock multipliers are dynamically enabled.

#02

CEDAR: Controlled and Event-Driven Demand Forecasting via Residual Decomposition

Aug 2026 arXiv:2608.25871v1 PDF

Authors & Institution

Junjie Meng, Ranxu Zhang, Zi-an Zhang, Chao Wang et al. (Alibaba 1688)

Target Modules

src/event_analyzer.py, src/models.py

Methodology Type

Two-Stage Decoupled Residual Decomposition

Key Theorem & Mathematical Formulation

$$\mathbf{s}_{t+1} = f_\theta(\mathbf{s}_{\le t}, \mathbf{a}_{\le t+1}) + \epsilon_t, \quad \hat{y}_{\mathrm{final}} = f_{\mathrm{quant}}(X) + \hat{\epsilon}_{\mathrm{event}}$$

Decouples baseline quantitative time-series transitions from qualitative event shock residuals ($\epsilon_t$). Stage I fits regularized Ridge regression on numerical commodity fundamentals; Stage II prompts Gemini 2.5 Flash for Tag Filtering & Regional Calendar Event Synthesis, outputting calibrated residual delta adjustments $\hat{\epsilon}_{\mathrm{event}}$.

#03

TraceBench: Controlled Evaluation of LLM Agents for Time-Series Root-Cause Attribution

Aug 2026 arXiv:2608.27182v1 PDF

Authors

Tommaso Bendinelli, Artur Dox, Christian Holz

Target Modules

src/event_analyzer.py, src/noaa_weather.py

Methodology Type

Structured Data Ingestion & Explicit Domain Prompts

Empirically demonstrated that time-series LLM agents explore data significantly better through compact numerical JSON/console formats than visual plots. Guided Midgley's token-efficient NOAA weather architecture (wxs.us 150-token summaries & 0-token deterministic SPC convective risk lookups).

#04

SAGE: Variate-Wise Semantic Augmentation for Vision-Language Time Series Forecasting

Aug 2026 arXiv:2608.26829v1 PDF

Authors

Haizhao Fan, Xinyi Le

Target Modules

src/feature_engineering.py

Methodology Type

Variate-Specific Domain Descriptors

Injects explicit variate-level metadata (physical units, autocorrelation decay rates, geographic hub profiles) into the feature engineering layer without requiring resource-heavy LLM inference inside critical real-time execution loops.

#05

Modeling Spatio-Temporal Locality in Multi-Step Forecasting of Geo-Referenced Time Series (SPALT)

Aug 2026 arXiv:2608.25698v1 PDF

Authors

Annunziata D'Aversa, Gianvito Pio, Michelangelo Ceci

Target Modules

src/locations/*/regional.py, src/metro_agents.py

Methodology Type

Spatio-Temporal Locality Trees

Guides Midgley's 8 regional calibration hubs (Tulsa OK, Newark DE, Cincinnati OH/KY, Greenville NC, Charlotte NC, Oakland CA, SF Bay Area, Port St. Lucie FL) to capture localized geographic autocorrelation while preserving specific state motor fuel excise tax and refining hub characteristics.

#06

LLM Agents for Time-Series: A Survey

Aug 2026 arXiv:2608.26226v1 PDF

Authors

Yilong Chen, Xiao Qin, Chenghao Liu, et al.

Target Documentation

AGENTS.md, docs/technical_breakdown.html

Methodology Type

Agentic Architecture Taxonomy

Provides the foundational taxonomy benchmarking Midgley's 8 specialized agents across Forecasting & Reasoning, Anomaly Detection & Diagnosis, and Decision Support / Counterfactual Scenario Simulation.

#07

Advances in Financial Machine Learning (Purged CPCV & PBO Overfitting Audit)

Wiley 2018 Wiley Text

Author

Marcos López de Prado

Target Modules

src/models.py, src/feature_auditor.py

Methodology Type

Purged & Combinatorial Purged Cross-Validation (CPCV)

Key Theorem & Mathematical Formulation

$$\text{Purge Condition: } i \in \text{Train} \iff [t_{i,\text{start}}, t_{i,\text{end}}] \cap [t_{j,\text{test,start}}, t_{j,\text{test,end}} + h_{\text{embargo}}] = \emptyset \quad \forall j \in \text{Test}, \quad \text{PBO} = \int_{-\infty}^0 f(\lambda) \, d\lambda$$

Purges overlapping forward label windows with post-test embargoes ($h_{\text{embargo}} = 5\text{ days}$) to eliminate temporal lookahead data leakage in 5-day horizon forecasting. Implements CSCV Probability of Backtest Overfitting (PBO) and Deflated Sharpe Ratio (DSR) in src/feature_auditor.py (Issue #146).

#08

A Decoder-Only Foundation Model for Time-Series Forecasting (TimesFM)

Oct 2023 / 2024 arXiv:2310.10688 PDF

Authors & Institution

Abhimanyu Das, Weihao Kong, Andrew Leach et al. (Google Research)

Target Modules

src/timesfm_forecaster.py

Methodology Type

Decoder-Only Foundation Model Forecasting

Wraps Google Research's decoder-only time-series foundation model with scikit-learn API compatibility, generating zero-shot point predictions and $P_{10}, P_{50}, P_{90}$ quantile uncertainty bounds alongside an analytical zero-shot fallback (Issues #185 & #112).

#09

Qlib: AI Investment Platform & RD-Agent: Autonomous Factor Mining (DDG-DA)

Sep 2020 / 2024 arXiv:2009.11189 PDF

Authors & Institution

Xiao Yang, Weiqing Liu, Dong Zhou et al. (Microsoft Research)

Target Modules

src/qlib_symbolic_engine.py, src/alpha_factor_miner.py, src/ddg_da_adapter.py

Methodology Type

Symbolic Alpha Mining & Dynamic Domain Adaptation

Key Theorem & Mathematical Formulation

$$IC_t = \text{Corr}(f_t, r_{t+h}), \quad IC_{IR} = \frac{\mu(IC)}{\sigma(IC)}, \quad w_i = \exp\left(-\gamma \|x_i - \bar{x}_{\text{recent}}\|^2\right)$$

Evaluates symbolic factor formulas for Information Coefficient ($IC > 0.03, IC_{IR} > 0.50$), prunes collinear factors ($|r| > 0.70$), and weights historical training samples using Gaussian RBF kernels across GMM market regimes to combat concept drift (Issue #127).

#10

Dynamic Volatility-Gated Persistence Blending (DV-GPB) & Empirical Residual Calibration

Sep 2026 Midgley Proprietary

Authors

Midgley Quantitative Research Group

Target Modules

src/models.py, src/dynamic_region.py

Methodology Type

Adaptive Sigmoid Gating & Empirical 95% Residual CI

Key Theorem & Mathematical Formulation

$$\lambda_{\text{vol}} = \frac{1}{1 + e^{-200.0 \cdot (\sigma_{14\text{d}} - 0.015)}}, \quad \hat{y}_{t+5} = \lambda_{\text{vol}} \hat{y}_{\text{model}} + (1 - \lambda_{\text{vol}}) y_t, \quad \text{CI}_{95\%} = \hat{y}_{t+5} \pm 1.96 \cdot \sigma_{\text{residual, 30d}}(r)$$

Transitions model predictions smoothly toward pure naive persistence during low-volatility plateaus ($\sigma_{14\text{d}} \ll 0.015$) while maintaining 100% responsiveness to real-world event shocks ($\sigma_{14\text{d}} > 0.015$), coupled with empirical 30-day residual standard error bounds ensuring $\ge 90\%$ coverage (Issue #214).

#11

CORE: A Global Aggregation of Open Access Research Papers

2012 / 2024 core.ac.uk

Authors

Petr Knoth, Zdenek Zdrahal

Target Modules

src/core_monitor.py

Methodology Type

Open-Access Literature Monitor

Integrates the CORE API v3 search works endpoint to monitor global open-access repositories for literature on refining crack spreads, energy commodity econometrics, and asymmetric retail price transmission during weekly review cycles (Issue #53).

#12

Sapient PRAXIST: Computer-Executable Autonomous Research Harness

Sep 2026 Midgley Framework

Authors

Midgley Quantitative Research Group

Target Modules

src/praxist_engine.py

Methodology Type

Empirical Hypothesis Testing & Automated Parameter Sweeps

Provides programmatic execution harnesses allowing LLM agents to formulate empirical feature hypotheses, execute out-of-sample backtests, compute paired $t$-tests and $p$-values, and optimize hyperparameters (Ridge $\alpha$, decay half-life $t_{1/2}$) autonomously (Issue #188).