This paper proposes Adaptive Event Driven Labeling (AEDL), a novel neurosymbolic framework that synthesizes unstructured semantic data with formal causal inference to disentangle simultaneous supply and demand shocks and introduces a Heterogeneous Multiagent Discussion (HAD) architecture.
Abstract
Identifiability in structural causal models remains a persistent challenge in high dimensional nonstationary environments where latent shocks are obscured by information asymmetries and regime shifts. Traditional econometric methods often rely on rigid recursive assumptions or sign restrictions that falter during periods of extreme volatility. This paper proposes Adaptive Event Driven Labeling (AEDL), a novel neurosymbolic framework that synthesizes unstructured semantic data with formal causal inference to disentangle simultaneous supply and demand shocks. We introduce a Heterogeneous Multiagent Discussion (HAD) architecture, wherein Large Language Model agents with distinct analytical personas engage in dialectic consensus protocols to generate continuous intensity weighted narrative instruments. To mitigate hallucination and enforce economic coherence, the framework integrates a symbolic verification layer and a Reflexion mechanism that iteratively updates causal priors based on posterior market deviations. Empirical validation on global energy market data from 2020 to 2025 demonstrates that the proposed framework significantly outperforms standard set identification techniques, achieving a supply shock identification F1 score of 0.89. Specifically, the model successfully decomposed the February 2021 Texas Freeze by identifying a supply shock magnitude of 0.8 and a demand shock magnitude of 0.7, which attributed 60.0% of the immediate price spike to supply constraints. Furthermore, the Reflexion mechanism corrected initial physical supply estimates for the 2022 Ukraine invasion from 0.9 down to 0.4, while accurately isolating a geopolitical uncertainty premium of 0.95. By rigorously bridging natural language reasoning with time series econometrics, this work advances the state of the art in automated causal discovery for macro scale complex systems.
Event-driven causal labeling in non-stationary time series is useful only when the system can state when its own root-cause decision is ambiguous. Existing neuro-symbolic multiagent pipelines improve semantic and structural consistency, but their confidence scores need not retain a stable error meaning after regime changes. This paper presents Conformal-AEDL, a resource-efficient extension that couples heterogeneous residual and propagation agents with reliability-weighted consensus, an explicit symbolic compliance layer, and online conformal prediction sets. The method outputs a set of plausible root causes rather than forcing a singleton, and adapts the conformal error budget when score surprises indicate rapid drift. Evaluation uses a nonlinear six-node structural causal model with four regimes and a semi-synthetic U.S. macroeconomic benchmark whose dynamics and residuals are estimated from 203 public quarterly observations while intervention labels remain controlled. Across ten seeds, Conformal-AEDL attains 0.929±0.022 and 0.917±0.029 coverage at a nominal 0.90 level, with average set sizes of 1.186 and 1.890. Relative to static split conformal prediction, this reduces set size by 26.0% and 29.5%, respectively. The results show that calibrated abstention can be added without inventing causal ground truth or relying on proprietary language-model calls, while also exposing an important limitation: immediate post-switch coverage remains difficult in the macro benchmark.
Elena Rossi, Lars Ø. Johansen· Academic Journal of Applied...· 0 citations
Structural causal models estimated on non-stationary time series are threatened by concept drift: the data-generating structure itself changes over time, silently invalidating a once-correct model. Recent neuro-symbolic pipelines such as Adaptive Event-Driven Labeling (AEDL) combine language-model reasoning with econometric identification, but they are trained one-shot and assume a fixed causal graph and a static anchor set. We propose DriftGuard-AEDL, a concept-drift-aware, continual extension that (i) estimates a contemporaneous causal graph online under a symbolic prior that encodes a known causal ordering and sign constraints; (ii) detects drift with a two-signal detector that couples a residual Page-Hinkley test with a lagged-reference test on the reduced-form coefficient stream; (iii) uses a symbolic verifier as an acceptance gate that admits an adaptation only when the re-estimated structure is economically coherent and materially changed; and (iv) adapts continually with anchor re-calibration and replay. On a controlled regime-switching structural VAR with known ground truth, DriftGuard-AEDL recovers the causal structure far better than drift-agnostic baselines (structural Hamming distance 1.9 vs. 7.3–8.2 and edge-F1 0.75 vs. ≤0.15) and matches a continuously retrained sliding-window model in predictive error while using 6 updates instead of 114; the symbolic gate cuts the false-alarm rate from 0.71 to 0.50 and the update count from 9.4 to 6.0. On the real Elec2 electricity-market benchmark (45,312 records), it recovers most of the adaptation gain of blind retraining at roughly one-fifth of the update cost and four times the speed. The results show that pairing lightweight symbolic constraints with drift-triggered continual learning yields causal inference that stays valid as the world changes.
Anika Shah, M. Greco, Freya Nielsen· Journal of Computing and Ele...· 0 citations
Financial time series forecasting tasks, specifically stock trend prediction and market attribution, are essential for quantitative investment and risk control. However, these tasks suffer from low signal-to-noise ratios and non-stationarity, stemming from the coupling of hierarchical drivers: market trends, sector rotations, and idiosyncratic dynamics. Existing methods, often relying on static sector labels or time-domain correlations, struggle to capture dynamic, multi-scale dependencies and lack the interpretability required for return attribution. To address this, we propose ResDIF, a Residual Disentanglement framework for Interpretable financial time series Forecasting inspired by Asset Pricing Theory (APT). ResDIF employs a progressive residual architecture via a Spectrally-Enhanced Temporal Encoding mechanism to explicitly decompose stock data into market, sector, and individual layers. Furthermore, a self-supervised orthogonal loss encourages feature separation, enabling the model to autonomously decouple systematic risks from idiosyncratic alpha. Beyond predictive modeling, ResDIF effectively unifies high predictive accuracy with granular structural attribution. By leveraging this intrinsic interpretability to quantify structural market risks, our framework further enables adaptive portfolio optimization through dynamic hedging and asset selection. Experiments demonstrate that ResDIF outperforms existing methods while providing actionable interpretability support for refined portfolio risk management.
Chengwei Fu, Gang Xiao, Yuchao Zhang et al.· Proceedings of the 32nd ACM...· 0 citations
A-SR, a self-evolving agentic framework that shifts the control unit from expression edits to role-conditioned evidence views, is proposed, a self-evolving agentic framework that shifts the control unit from expression edits to role-conditioned evidence views.
Wenxiao Zhao, Dong Liu, Kaiyi Xu et al.· 2 citations
Time series forecasting models are widely used in high-stakes settings, yet their predictions remain difficult to interpret because existing post-hoc methods often ignore temporal dependence and fail to provide horizon-specific explanations. We propose a model-agnostic explainability framework that explains forecasting predictions by attributing each forecast horizon to temporally relevant historical lags. The framework models forecasting as a latent trajectory and introduces semantic flow to quantify how information evolves across time in the model's internal representations. By aggregating semantic flow, it constructs a lag-horizon attribution matrix that captures horizon-resolved temporal influence. To improve explainability, we further generate structure-preserving perturbations and fit sparse local surrogate models, producing human-readable and temporally coherent explanations. We evaluate the method using faithfulness and stability diagnostics across multiple benchmark datasets. Results show that the semantic-flow variant achieves competitive or superior faithfulness compared to standard post-hoc baselines, while being substantially more computationally efficient. Stability analysis further demonstrates that the explanations are robust and identifies regimes where interpretation should be applied with caution.
