Model
The underlying AI or statistical capability: an LLM, tree model, time-series model, neural network, graph model, anomaly detector, or other algorithmic system.
Models
Models are classified, versioned, and governed before they become forecasts
iPulse uses frontier language models for deep reasoning and narrative market intelligence, while the wider architecture also supports supervised, unsupervised, time-series, neural, graph, anomaly, and simulation model families for signals and validation.
Updated July 3, 2026
Overview
A model response becomes useful when it is structured
iPulse does not treat a raw model answer as the finished product. A model runs inside a controlled task with a defined AI Agent, execution mode, prompt assembly, input format, output schema, forecast horizon, and validation path. That structure is what turns a model call into a forecast record that can be inspected later.
The public product currently emphasizes AI Agents with deep reasoning and readable investment narratives: users can compare thesis, risks, scenarios, and disagreement. Under the hood, the platform is designed to support many other model families as signal builders, validators, feature engines, and future forecast layers.
Agent
The analytical identity and investment framework that tells a capable model what lens to apply to market evidence.
Task
The configured assignment: mode, data inputs, prompt assembly, output schema, horizon, validation requirements, and lineage fields.
Frontier LLMs
The visible research layer uses frontier language models
The investment thesis experience references several frontier model providers because iPulse is built for model comparison and portability. These models are especially valuable when the task needs long-context synthesis, reasoning over messy evidence, structured JSON output, and human-readable explanation.
Gemini
Deep reasoning, long-context synthesis, multimodal understanding, structured market narratives.
Claude
Anthropic
Long-form reasoning, careful synthesis, report writing, and nuanced risk explanation.
ChatGPT
OpenAI
General reasoning, coding, structured analysis, tool use, and natural-language explanation.
Grok
xAI
Fast narrative analysis, market discourse awareness, and alternative model comparison.
The provider logo is not the methodology. The methodology is the controlled wrapper around the model: AI Agent identity, mode, prompt assembly, context files, output schema, validation, consensus packaging, and lineage.
Classification
The model registry starts with a broad AI taxonomy
Shared-base enums classify models across learning paradigms, problem types, algorithms, architecture structures, outputs, metrics, and frameworks. This lets iPulse describe a Gemini-style foundation model and a LightGBM tabular model inside the same governance vocabulary without pretending they are the same thing.
Model taxonomy
How iPulse classifies models before they become signals
Learning Paradigm
SupervisedUnsupervisedSemi-supervisedReinforcementTransfer learningActive learning
Problem Type
RegressionClassificationRankingClusteringAnomaly detectionTime-series forecastingGenerationQA
Algorithm Family
Linear/GLMTrees and boostingARIMA/state-spaceCNN/RNN/LSTM/GRUTransformersGNNsAutoencodersDiffusion
Architecture
Single modelSequentialCascadeEnsembleStacked ensembleVoting ensembleBoostingMixture of experts
Outputs
Regression valueClass probabilityTime-series forecastRankingEmbeddingAnomaly scoreFeature importanceExplanation
Frameworks
scikit-learn, TensorFlow, PyTorch, Keras, XGBoost, LightGBM, CatBoost, statsmodels, Prophet, Darts, sktime, tsfresh, PyOD, Hugging Face.
Finance uses
Forecasting, ranking, anomaly detection, regime clustering, factor modeling, event detection, feature extraction, and narrative synthesis.
Governance
Specifications, versions, serving instances, validation metrics, approval notes, performance scores, and retirement records stay separate.
Linear and statistical models
Examples include linear regression, logistic regression, ridge, lasso, elastic net, ARIMA, SARIMA, exponential smoothing, and state-space models. In finance, they are useful for baselines, factor sensitivity, return decomposition, and transparent sanity checks.
Tree and boosting models
Examples include random forests, XGBoost, LightGBM, and CatBoost. These are strong for tabular finance data such as fundamentals, valuation ratios, credit features, event flags, and ranking signals.
Neural sequence models
Examples include RNNs, LSTMs, GRUs, temporal convolutional networks, N-BEATS, and transformers. These can model non-linear temporal patterns across prices, indicators, sentiment, flows, and macro sequences.
CNNs and multidimensional detection
CNN families such as ResNet-style or EfficientNet-style architectures can be useful when the input behaves like an image or grid: limit-order-book tensors, volatility heatmaps, chart-state images, satellite data, or sector correlation maps.
Graph models
Graph neural networks can represent relationships: supplier networks, ownership structures, board links, capital flows, sector dependencies, sanctions exposure, or customer/vendor concentration.
Foundation and language models
Large language and multimodal models help synthesize filings, news, macro context, event records, and advisor-style reasoning into structured narratives, risks, scenarios, and forecast explanations.
Lineage
Model spec, model version, and deployed instance are separate
This separation is central to transparency. A model specification describes the family. A model version describes the exact capability snapshot available at a point in time. A serving instance describes the endpoint or execution environment that actually handled a request.
Example model lineage
A Gemini-style model record is split into three different objects
01 Model Spec
Family identity
Example name: google_gemini_2_5_pro. Defines provider, source, learning paradigm, problem types, algorithms, architecture structure, strengths, weaknesses, and recommended use cases.
02 Model Version
Capability snapshot
Example name: gemini_2_5_pro_20250626. Defines API identifier, release date, knowledge cutoff, input and output modalities, context limits, supported modes, version status, release notes, and known limitations.
03 Serving Instance
Callable endpoint
Example name: gemini_2_5_pro_genai_client. Defines hosting provider, region, endpoint, authentication reference, rate limits, monitoring, traffic split, health checks, and production approval.
Model specification
The blueprint: provider, model family, learning paradigm, problem types, algorithm categories, architecture structure, general strengths, weaknesses, and intended use cases.
Model version
The immutable capability record: API identifier, release date, knowledge cutoff, input/output modalities, context limits, supported modes, known limitations, and retirement status.
Serving instance
The callable service: endpoint, hosting provider, region, auth method, monitoring, rate limits, traffic percentage, scaling behavior, and approval metadata.
What improved
Model governance exists because model names alone are not enough
A public provider model name is too ambiguous for serious forecast review. The same family can receive new API identifiers, capability changes, context limits, tool settings, knowledge-cutoff changes, serving regions, and retirement states. iPulse separates those layers so a future review can ask which capability was actually used, not only which brand was displayed.
Version changes become reviewable
When a model version changes, iPulse can compare forecast behavior before and after the change instead of mixing two capabilities under one vague label.
Serving changes stay operational
Endpoint, region, monitoring, rate limits, traffic allocation, and approval metadata belong to serving instances so operational changes do not rewrite model identity.
Non-LLM models fit the same registry
Tree models, time-series forecasters, anomaly detectors, neural networks, and graph models can be governed alongside LLMs when each family is classified explicitly.
Config Details can stay readable
Users can see the meaningful model lineage in product surfaces while low-level admin-only deployment details remain internal where appropriate.
Supervised Learning
Supervised models turn labeled examples into numeric signals
Supervised models learn from examples where the target is known. In finance, that might mean predicting next period return, classifying drawdown risk, ranking assets by expected quality, or estimating whether an event is likely to matter for price.
Regression and classification
Linear regression, logistic regression, ridge, lasso, elastic net, SVMs, and neural MLPs can support transparent baselines, factor studies, probability estimates, and model sanity checks.
Tree and boosting libraries
Random forests, XGBoost, LightGBM, and CatBoost are often strong on tabular data such as fundamentals, valuation ratios, event flags, quality metrics, and macro indicators.
Unsupervised Learning
Unsupervised models find structure before a label exists
Unsupervised methods help discover patterns that are not already labeled. For markets, that can mean regime clustering, anomaly detection, peer grouping, factor compression, or identifying unusual combinations of valuation, sentiment, volatility, and macro stress.
Clustering and dimensionality reduction
K-means, hierarchical clustering, DBSCAN, Gaussian mixtures, PCA, t-SNE, UMAP, and LDA can help map regimes, peer groups, and compressed factor spaces.
Anomaly detection
Isolation Forest, one-class SVM, LOF, and autoencoders can help flag strange trading behavior, accounting outliers, volatility shocks, or data quality breaks.
Time Series
Time-series models specialize in sequence, seasonality, and horizon
Time-series forecasting is different from ordinary tabular prediction because order matters. Prices, volumes, rates, earnings revisions, inflation, and sentiment all evolve through time, and the model must respect horizon, lag, regime, autocorrelation, volatility, and look-ahead risk.
Statistical and neural forecasters
AR, ARIMA, SARIMA, exponential smoothing, state-space models, Prophet, N-BEATS, temporal convolutional networks, and transformer forecasters can each serve different horizon and interpretability needs.
Time-series foundation models
Google Research TimeFM is a notable decoder-only foundation model for time-series forecasting. iPulse Labs can evaluate this family for zero-shot and few-shot forecasting signals alongside more traditional models.
In the public product today, these models are best understood as signal and validation candidates rather than the main visible experience. The visible experience remains the AI Agent narrative layer, where users can read why a forecast was produced and compare disagreements across agents.
Neural And Multimodal
Neural models can detect patterns across text, images, graphs, and tensors
Not all market evidence is a clean table. Some useful signals arrive as documents, transcripts, chart images, heatmaps, maps, satellite imagery, audio, video, or relationship graphs. The model taxonomy already leaves room for CNNs, RNNs, LSTMs, GRUs, transformers, GNNs, VAEs, GANs, and diffusion-style models.
CNN-style detection
CNNs can help when financial evidence becomes spatial or grid-like: chart-state images, order-book tensors, volatility surfaces, sector correlation heatmaps, or satellite-derived activity signals.
Transformers and graph models
Transformers help with long sequences and multimodal reasoning. Graph models can represent ownership, suppliers, customers, political relationships, capital flows, or sector dependencies.
Labs Roadmap
iPulse Labs explores multiple model families, not only LLMs
At this stage, iPulse mostly exposes AI Agents with deep reasoning capabilities because users need readable narratives, risks, scenarios, and transparent disagreement. But the architecture is intentionally broader: supervised models, unsupervised models, time-series forecasters, anomaly detectors, graph models, and multimodal systems can all become features, validators, or specialist signals.
The product direction is hybrid intelligence: language models explain and synthesize; specialized models can detect, rank, forecast, cluster, and validate. The more these layers are separated in the registry, the easier it becomes to compare performance and retire weak configurations without hiding the historical record.
Config Details
Where users see model-related details
Eligible signed-in users can use Config Details to inspect readable model and task configuration context for a specific prediction. Access depends on the subscription plan and whether the plan unlocks the relevant asset, batch, or advisor report. Some low-level provider, endpoint, or admin-only deployment details may stay internal even when the public methodology explains the architecture.