TimeSeriesScientist: Agentic Forecasting Framework

• TSci is a general, domain-agnostic forecasting framework that decomposes the workflow into four specialized agents while minimizing human intervention.
• It employs a staged decision process where Curator, Planner, Forecaster, and Reporter collaboratively manage diagnostics, model tuning, ensembling, and report synthesis.
• Experimental evaluations demonstrate TSci reduces errors by up to 38.2% over LLM and statistical baselines on diverse datasets, proving its practical efficacy.

TimeSeriesScientist, abbreviated TSci, is an LLM-driven agentic framework for general time series forecasting that treats forecasting as a staged decision process rather than as a single model-fitting step. It is designed for settings with thousands of short, noisy series that vary in frequency, quality, and forecast horizon, where the dominant cost lies in preprocessing, validation, and ensembling rather than in fitting an individual predictor. The framework consists of four specialized agents—Curator, Planner, Forecaster, and Reporter—and produces both forecasts and a comprehensive, transparent report with natural-language rationales (Zhao et al., 2 Oct 2025).

1. Definition, scope, and problem setting

TSci addresses univariate forecasting from raw input data $\mathcal{D}=\{x_t\}_{t=1}^T$. Its stated objective is a general, domain-agnostic framework that minimizes human intervention while remaining interpretable and extensible across tasks. In this formulation, the forecasting workflow is decomposed into diagnosis and preprocessing, model-pool selection and tuning, adaptive model selection and ensembling, and report synthesis (Zhao et al., 2 Oct 2025).

A central design claim is that prevailing statistical and deep learning models are often tailored to specific datasets or domains and generalize poorly, whereas forecasters in practice must manage heterogeneous operational burdens before any final prediction is produced. TSci therefore positions the main unit of automation not as a single predictor but as an end-to-end workflow controller. A common misconception is to equate TSci with a new forecasting architecture analogous to a Transformer or a state-space model; the framework is instead a tool-augmented multi-agent system that orchestrates diagnostics, model selection, validation, and reporting (Zhao et al., 2 Oct 2025).

Within the broader time-series software landscape, this places TSci alongside, but conceptually above, systems that solve narrower layers of the stack. The tsibble framework formalizes temporal data with an explicit index and key for tidy pipelines (1901.10257), cesium provides an end-to-end platform for featurization and machine learning with notebook export (Naul et al., 2016), and GluonTS supplies probabilistic forecasting abstractions such as Estimator, Predictor, Forecast, and Evaluator (Alexandrov et al., 2019). This suggests that TSci’s distinctive contribution is not a new base forecaster alone, but an agentic orchestration layer over the forecasting lifecycle.

2. Four-agent architecture and staged workflow

TSci organizes forecasting through four specialized agents: Curator, Planner, Forecaster, and Reporter. At a high level, the pipeline is described as follows: raw time series $\mathcal{D}$ enter Curator; Curator produces cleaned data, visualizations, and analysis; Planner narrows the model hypothesis space and tunes candidates; Forecaster fits, validates, and selects a final model configuration or ensemble strategy; Reporter synthesizes the process into a human-readable report (Zhao et al., 2 Oct 2025).

The workflow can be written as a sequence of transformations: $$(Q,V,A)\gets \mathrm{Curator}(\mathcal{D}), \qquad \mathcal{M}_p\gets \mathrm{Planner.SelectModels}(Q,V,A),$$

$$(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}})\gets \mathrm{Planner.Tune}(\mathcal{M}_p,\tilde{\mathcal{D}}_{\mathrm{val}}),$$

$$\hat y_{t+1:t+H}\gets \mathrm{Forecaster}(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}}), \qquad \mathcal{R}\gets \mathrm{Reporter}(Q,V,A,\mathcal{M}_{\mathrm{selected}},\hat y).$$

This decomposition is operational rather than merely descriptive. Curator outputs diagnostic artifacts, Planner outputs a selected subset of models and their tuned configurations, Forecaster outputs final predictions, and Reporter outputs a structured report. Because each stage exposes intermediate products, the framework is described as a white-box system with transparent natural-language rationales and comprehensive reports (Zhao et al., 2 Oct 2025).

The architecture also makes explicit that TSci separates representation of evidence from forecast generation. Diagnostics $Q$, visualizations $V$, and temporal-structure analysis $A$ are produced before model selection, rather than being implicit side products of a learned end-to-end network. This suggests an emphasis on procedural interpretability rather than solely post hoc explanation.

3. Curator and Planner

The Curator agent performs three steps: Quality Diagnostics, Visualization Generation, and Temporal Structure Profiling. It computes

$Q=\mathcal{A}_f(\mathcal{D})=(S,M,O,\pi),$

where $$S=(\mu,\sigma,x_{\min},x_{\max},\tau_{\mathrm{trend}})$$, $\mathcal{D}$0 and $\mathcal{D}$1 record missing-value and outlier indices, and $\mathcal{D}$2 is the chosen preprocessing strategy; it then applies a transformation $\mathcal{D}$3 to produce cleaned data $\mathcal{D}$4 (Zhao et al., 2 Oct 2025).

Outlier detection in Curator includes three explicit schemes. Under Rolling IQR, $\mathcal{D}$5 is an outlier if it falls below $\mathcal{D}$6 or above $\mathcal{D}$7 on a rolling window. Under Rolling Z-Score, $\mathcal{D}$8 is an outlier if $\mathcal{D}$9 with $$(Q,V,A)\gets \mathrm{Curator}(\mathcal{D}), \qquad \mathcal{M}_p\gets \mathrm{Planner.SelectModels}(Q,V,A),$$0. Under the Percentile Rule, an observation is flagged when it lies below $$(Q,V,A)\gets \mathrm{Curator}(\mathcal{D}), \qquad \mathcal{M}_p\gets \mathrm{Planner.SelectModels}(Q,V,A),$$1 or above $$(Q,V,A)\gets \mathrm{Curator}(\mathcal{D}), \qquad \mathcal{M}_p\gets \mathrm{Planner.SelectModels}(Q,V,A),$$2. The paper states that outlier handling includes clipping, interpolation, forward/backward fill, and smoothing (Zhao et al., 2 Oct 2025).

For stationarity assessment, Curator may call external tools for the Augmented Dickey–Fuller (ADF) test and the KPSS test. It also generates three plot types through LLM-driven prompts with Matplotlib and seaborn: a time-series overview with rolling mean and standard deviation, a seasonal decomposition $$(Q,V,A)\gets \mathrm{Curator}(\mathcal{D}), \qquad \mathcal{M}_p\gets \mathrm{Planner.SelectModels}(Q,V,A),$$3, and ACF/PACF plots up to lag 40. From cleaned data $$(Q,V,A)\gets \mathrm{Curator}(\mathcal{D}), \qquad \mathcal{M}_p\gets \mathrm{Planner.SelectModels}(Q,V,A),$$4 and visualizations $$(Q,V,A)\gets \mathrm{Curator}(\mathcal{D}), \qquad \mathcal{M}_p\gets \mathrm{Planner.SelectModels}(Q,V,A),$$5, it prompts the LLM to extract

$$(Q,V,A)\gets \mathrm{Curator}(\mathcal{D}), \qquad \mathcal{M}_p\gets \mathrm{Planner.SelectModels}(Q,V,A),$$6

where $$(Q,V,A)\gets \mathrm{Curator}(\mathcal{D}), \qquad \mathcal{M}_p\gets \mathrm{Planner.SelectModels}(Q,V,A),$$7 is trend type, $$(Q,V,A)\gets \mathrm{Curator}(\mathcal{D}), \qquad \mathcal{M}_p\gets \mathrm{Planner.SelectModels}(Q,V,A),$$8 seasonality strength, and $$(Q,V,A)\gets \mathrm{Curator}(\mathcal{D}), \qquad \mathcal{M}_p\gets \mathrm{Planner.SelectModels}(Q,V,A),$$9 the stationarity verdict (Zhao et al., 2 Oct 2025).

The Planner then narrows the model hypothesis space in three stages: model selection, hyperparameter optimization, and model ranking. It maps multimodal diagnostics $$(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}})\gets \mathrm{Planner.Tune}(\mathcal{M}_p,\tilde{\mathcal{D}}_{\mathrm{val}}),$$0 to a candidate pool

$$(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}})\gets \mathrm{Planner.Tune}(\mathcal{M}_p,\tilde{\mathcal{D}}_{\mathrm{val}}),$$1

where $$(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}})\gets \mathrm{Planner.Tune}(\mathcal{M}_p,\tilde{\mathcal{D}}_{\mathrm{val}}),$$2 is a predefined library of 21 models, including ARIMA, ETS, Prophet, XGBoost, LSTM, Transformer, etc. For each candidate $$(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}})\gets \mathrm{Planner.Tune}(\mathcal{M}_p,\tilde{\mathcal{D}}_{\mathrm{val}}),$$3, Planner samples $$(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}})\gets \mathrm{Planner.Tune}(\mathcal{M}_p,\tilde{\mathcal{D}}_{\mathrm{val}}),$$4 of size $$(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}})\gets \mathrm{Planner.Tune}(\mathcal{M}_p,\tilde{\mathcal{D}}_{\mathrm{val}}),$$5 and minimizes validation MAPE: $$(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}})\gets \mathrm{Planner.Tune}(\mathcal{M}_p,\tilde{\mathcal{D}}_{\mathrm{val}}),$$6 It then ranks tuned candidates by validation MAPE and keeps the top $$(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}})\gets \mathrm{Planner.Tune}(\mathcal{M}_p,\tilde{\mathcal{D}}_{\mathrm{val}}),$$7 models as

$$(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}})\gets \mathrm{Planner.Tune}(\mathcal{M}_p,\tilde{\mathcal{D}}_{\mathrm{val}}),$$8

This design explicitly couples diagnosis-driven model narrowing with metric-driven tuning rather than performing blind model sweep over the full library (Zhao et al., 2 Oct 2025).

4. Forecaster and Reporter

The Forecaster agent takes the selected models $$(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}})\gets \mathrm{Planner.Tune}(\mathcal{M}_p,\tilde{\mathcal{D}}_{\mathrm{val}}),$$9 and their validation metrics $$\hat y_{t+1:t+H}\gets \mathrm{Forecaster}(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}}), \qquad \mathcal{R}\gets \mathrm{Reporter}(Q,V,A,\mathcal{M}_{\mathrm{selected}},\hat y).$$0, aggregates multiple metrics $$\hat y_{t+1:t+H}\gets \mathrm{Forecaster}(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}}), \qquad \mathcal{R}\gets \mathrm{Reporter}(Q,V,A,\mathcal{M}_{\mathrm{selected}},\hat y).$$1 into a scalar score,

$$\hat y_{t+1:t+H}\gets \mathrm{Forecaster}(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}}), \qquad \mathcal{R}\gets \mathrm{Reporter}(Q,V,A,\mathcal{M}_{\mathrm{selected}},\hat y).$$2

and chooses an ensemble strategy (Zhao et al., 2 Oct 2025).

Three ensemble strategies are specified. Under Single-Best, the chosen model is

$$\hat y_{t+1:t+H}\gets \mathrm{Forecaster}(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}}), \qquad \mathcal{R}\gets \mathrm{Reporter}(Q,V,A,\mathcal{M}_{\mathrm{selected}},\hat y).$$3

Under Performance-Aware Averaging, the weights are defined through

$$\hat y_{t+1:t+H}\gets \mathrm{Forecaster}(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}}), \qquad \mathcal{R}\gets \mathrm{Reporter}(Q,V,A,\mathcal{M}_{\mathrm{selected}},\hat y).$$4

followed by clipping and shrinkage toward uniform weighting. Under Robust Aggregation, TSci uses the median or a trimmed mean across model forecasts. Final evaluation reports test MAE and RMSE (Zhao et al., 2 Oct 2025).

The Reporter agent synthesizes a comprehensive report $$\hat y_{t+1:t+H}\gets \mathrm{Forecaster}(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}}), \qquad \mathcal{R}\gets \mathrm{Reporter}(Q,V,A,\mathcal{M}_{\mathrm{selected}},\hat y).$$5 with five sections:

1. ensemble forecast $$\hat y_{t+1:t+H}\gets \mathrm{Forecaster}(\mathcal{M}_{\mathrm{selected}},S_{\mathrm{val}}), \qquad \mathcal{R}\gets \mathrm{Reporter}(Q,V,A,\mathcal{M}_{\mathrm{selected}},\hat y).$$6 with confidence intervals;
2. performance summary for each model and the ensemble;
3. interpretability report covering model selection, hyperparameter choices, ensemble weights, and forecast confidence and assumptions;
4. visualization suite for preprocessing and forecast plots;
5. workflow documentation as a decision-provenance log (Zhao et al., 2 Oct 2025).

This reporting layer is a defining feature of TSci. Earlier systems expose reproducibility through data structures, notebooks, or evaluation APIs—such as tsibble’s explicit temporal metadata (1901.10257) or cesium’s JSON-logged notebook export (Naul et al., 2016)—but TSci makes report synthesis itself a first-class component of the forecasting system. This suggests that interpretability in TSci is procedural and documentary as much as statistical.

5. Experimental evaluation

TSci is evaluated on eight public benchmarks covering five domains. The datasets are listed with point counts and frequencies in the paper’s evaluation summary (Zhao et al., 2 Oct 2025).

Dataset(s)	Frequency	#points
ETTh1, ETTh2	1 h	17,420
ETTm1, ETTm2	15 m	69,680
Weather	10 m	52,696
Electricity (ECL)	1 h	26,304
Exchange	1 d	7,588
ILI	1 w	966

The study reports MAE and MAPE and compares TSci against both statistical baselines—ARIMA, ETS, TBATS—and LLM-based baselines—GPT-4o, Gemini-2.5 Flash, Qwen-Plus, DeepSeek-v3, and Claude-3.7. The headline result is that TSci reduces error by 10.4% versus statistical methods and 38.2% versus LLM baselines, with full results by dataset and horizon reported in the appendix (Zhao et al., 2 Oct 2025).

The benchmark mix is operationally significant. ETTh/ETTm and Weather provide higher-frequency settings, Exchange provides daily data, and ILI provides weekly data. This suggests that the framework is being evaluated across heterogeneous sampling regimes rather than only on a single industrial benchmark family.

6. Relation to adjacent methods, misconceptions, and limitations

TSci belongs to a broader line of systems that aim to reduce the engineering burden of time-series analysis, but its emphasis differs from feature-extraction platforms and probabilistic forecasting libraries. hctsa automates massive feature extraction over approximately 7,700 time-series features and then applies univariate or multivariate selection and modeling (Fulcher et al., 2016). cesium provides featurization, model building, prediction, and reproducible notebook export through a Python library and web interface (Naul et al., 2016). GluonTS standardizes probabilistic model development, training, prediction, and evaluation for forecasting workloads (Alexandrov et al., 2019). By contrast, TSci frames the entire workflow as an LLM-orchestrated sequence of diagnostic, selection, ensembling, and reporting decisions (Zhao et al., 2 Oct 2025).

A second common misconception is that forecasting performance alone exhausts the contribution. The paper instead stresses interpretability, transparency, and extensibility: each agent logs clear rationales, the final report is fully transparent, and new diagnostics, models, or ensemble strategies can be plugged into the framework. This suggests that TSci should be understood as a forecasting operations framework as much as a benchmarked predictor (Zhao et al., 2 Oct 2025).

The paper also states three principal limitations and future directions. First, efficiency remains constrained by LLM-call latency, making batching and caching open issues. Second, extending Curator’s diagnostics to multivariate or multimodal series is identified as a challenge. Third, the framework leaves open knowledge infusion, including integration of exogenous signals or external knowledge bases (Zhao et al., 2 Oct 2025).

These limitations delineate the current scope of TSci. It is presented as a general-purpose AI agent for forecasting, but its published formulation centers on univariate series and on workflow automation around a predefined model library. Its contribution therefore lies in turning preprocessing, model selection, validation, ensembling, and reporting into an explicit, inspectable agent pipeline rather than in replacing the underlying forecasting methods with a single new architecture (Zhao et al., 2 Oct 2025).