From Raw Data to Actionable Forecasts in Under Five Minutes
SIA, built by ResEt AI, addresses a critical gap in enterprise demand and operations planning: most organizations spend more time wrestling data and tuning models than extracting business value from forecasts.
SIA unifies two purpose-built engines into a single, self-tuning pipeline that works across hourly, shift, daily, weekly, biweekly, and monthly frequencies: a Unified Forecasting System for generating production-quality predictions, and a Forecastability Engine for assessing which items are worth forecasting at all.
The platform automatically ingests raw data, infers schema and frequency, detects outliers and structural patterns, trains and blends multiple model families, applies safety guardrails, and delivers richly formatted interactive reports, all with minimal configuration. An optional AI Vision layer adds chart-based reasoning to guide model selection and parameter tuning.
SIA's demand forecasting and forecastability capabilities were developed in partnership with Addverb, a global robotics and warehouse automation company. Addverb provides the operational data, deployment infrastructure, and domain expertise that ground these capabilities in real-world logistics and manufacturing conditions. Addverb's Warehouse Execution System, Concinity, serves as both the source of the demand signal SIA trains on and the execution layer where forecasts drive action.
Complementary Strengths: Intelligence Meets Operations
SIA's forecasting and forecastability capabilities are the result of a collaboration between two companies that bring complementary and non-overlapping strengths to the manufacturing and supply chain intelligence space.
ResEt AI
AI agents for manufacturing growth. ResEt AI builds the forecasting algorithms, the forecastability scoring framework, and the AI intelligence layer that power SIA's demand forecasting and forecastability capabilities. SIA is ResEt AI's broader decision intelligence platform, which translates complex operational data into plain-language insights so every stakeholder, from the shop floor to the boardroom, can act with clarity.
reset-ai.com →Addverb
A global robotics company specializing in warehouse automation across five continents. Addverb's ASRS, AMR, and conveyor deployments are where the demand signal lives. Their Warehouse Execution System, Concinity, holds the task history, inventory movement records, inbound/outbound order flows, and shift-level throughput data that SIA needs to train and validate forecasts. Addverb's deep domain expertise in shift structures, equipment downtime patterns, and seasonal throughput cycles directly informs SIA's physics-informed modeling layer.
addverb.com →Why this combination works: ResEt AI brings the algorithmic intelligence. Addverb brings the operational data, deployment infrastructure, and domain knowledge that grounds every forecast in physical reality. Together, they close the loop between prediction and action: SIA's forecasts flowing through Concinity into the same warehouse operations that generated the training data.
Why Enterprise Demand Forecasting Still Fails
Enterprise demand forecasting is deceptively hard. Not because individual algorithms are complex, but because the end-to-end pipeline has dozens of failure modes that silently degrade accuracy.
Data Chaos
Real-world data arrives in inconsistent formats, mixed encodings, ambiguous date columns, and varying compression formats. Manual preprocessing is error-prone and time-consuming.
Frequency Mismatch
Business decisions happen at specific cadences (shift-level, weekly, monthly) but data often arrives at a different granularity. Naive aggregation destroys signal.
One-Size-Fits-None
A single algorithm rarely wins across all series. Intermittent demand, trend shifts, and seasonal patterns each favor different modeling approaches.
Blind Spot: Forecastability
Teams waste effort forecasting items with insufficient or erratic data. Without a pre-assessment, bad forecasts propagate into inventory and planning systems undetected.
The operational reality: In warehouse environments like those Addverb operates, these problems compound. Shift-level throughput data from ASRS systems, AMR fleets, and conveyor lines arrives at mixed frequencies, with structural zeros from facility closures and equipment downtime baked in. SIA was designed to handle exactly this kind of operational data natively.
Dual-Engine Design with Shared Data Infrastructure
SIA is built around two complementary engines that share common data infrastructure and can operate independently or in concert.
When deployed alongside Addverb infrastructure, data flows in from Concinity (task history, inventory movements, order flows, shift throughput) and forecast outputs flow back into Concinity to drive operational decisions. The platform also accepts data directly from ERP systems, MES, SCADA, or flat CSV files for environments outside the Addverb ecosystem.
From Ingestion to Prediction, with Intelligent Automation at Every Stage
5.1 Intelligent Data Preprocessing
SIA begins with robust ingestion that handles compressed files (gzip, bz2, zip, xz), encoding fallbacks, and ambiguous delimiters. An AI-powered analysis step automatically infers the correct date column, target column, and data frequency from the file header and sample rows, eliminating the need for manual schema specification.
Once loaded, the preprocessing layer handles frequency normalization across seven supported tiers, from hourly and 8-hour shift data through to monthly aggregates. A purpose-built shift-bucketing algorithm correctly bins sub-daily data into operational shift windows, accounting for configurable shift start hours and cross-midnight boundaries.
Structural Pattern Detection: SIA automatically identifies structural off-days (e.g., facilities closed on weekends), structural off-hours, and seasonal calendar patterns. These structural zeros are distinguished from missing data, a critical distinction that prevents imputation from injecting false demand signal. In Addverb-operated warehouses, this means SIA correctly handles planned downtime, shift changeovers, and facility-level closures that Concinity records as part of normal operations.
5.2 Adaptive Outlier Treatment
Outlier detection adjusts its sensitivity by data frequency. Higher-frequency data (hourly, shift) uses tighter bounds, while lower-frequency data (monthly) applies wider tolerances to avoid over-smoothing limited observations. The engine employs a cascading strategy (starting with MAD-based detection and falling back to percentile methods for small samples) with configurable floor multipliers to prevent false positives.
5.3 Multi-Family Model Training
Rather than relying on a single algorithm, SIA trains multiple model families in parallel and selects the best combination based on validated performance:
Machine Learning Track
Exponential Smoothing (ETS) with auto-selected seasonality, SARIMAX with automated order selection and stationarity testing, and Gradient Boosting with lagged features. Each is independently tuned via a comprehensive auto-tuning engine.
Physics-Informed Track
A Fourier-decomposition model captures trend, multiple seasonal harmonics, and calendar effects using regularized regression. Includes adaptive transform selection, residual AR dynamics, and MAPE-weighted loss. The structural rules this model depends on (shift patterns, equipment downtime cycles, seasonal throughput variation) are precisely the operational context that Addverb understands and Concinity records.
5.4 Hybrid Blending & Ensemble
The blending layer finds the optimal convex combination of ML and Physics predictions through constrained optimization. SIA evaluates multiple blending strategies, including NNLS-based weighting and scale-bias correction, and automatically selects the approach that minimizes the priority accuracy metric on held-out validation data.
An auto-switch mechanism detects when one model family significantly outperforms the other and cleanly switches to single-model output, avoiding the dilution that forced ensembling can introduce.
5.5 Comprehensive Auto-Tuning
The auto-tuning engine operates across all model families with three strategy modes (fast, adaptive, and comprehensive) automatically selected based on dataset size and complexity. It performs hyperparameter search with cross-validated scoring, learning curve analysis, and feature importance profiling. The tuning metric is intelligently chosen based on data characteristics: wMAPE for data with significant zeros, MAPE for clean continuous series, and sMAPE as a balanced alternative.
5.6 Multi-Frequency Disaggregation
When the optimal modeling frequency differs from the business's requested frequency, SIA performs intelligent disaggregation. For example, a model trained at weekly granularity can produce daily or shift-level forecasts by applying learned intra-period profiles (weighted day-of-week distributions, intra-day patterns, and seasonal shape adjustments) that preserve the aggregate forecast total while distributing it realistically.
Supported Frequencies
| Frequency | Typical Use Case |
|---|---|
| Hourly | Energy trading, grid balancing |
| Shift (8-hour) | Manufacturing, warehouse operations, Concinity task scheduling |
| Daily | Retail demand, logistics, fulfillment |
| Weekly | Replenishment planning |
| Biweekly | Payroll-driven demand |
| Monthly | S&OP, budgeting, capacity planning |
Know What to Forecast Before You Forecast It
Before committing resources to forecasting, organizations need to know which items are worth forecasting. SIA's Forecastability Engine answers this question with a rigorous, multi-dimensional scoring framework.
6.1 The Scoring Framework
Every item in the dataset is evaluated through a structured waterfall that first applies hard data-sufficiency gates, then computes a composite forecastability score from 0 to 100 based on four weighted dimensions:
Data Sufficiency evaluates whether the item has enough history, low missing-data ratios, and sufficient non-zero observations to support reliable modeling. Intermittency measures how frequently demand occurs. High zero-demand ratios and large average inter-demand intervals reduce the score. Variability captures demand volatility through coefficient of variation and spike analysis. Signal Strength assesses whether the time series contains learnable patterns: autocorrelation, seasonal decomposition strength, and proxy backtest performance.
6.2 Gate-Based Triage
Before the formula score is computed, a set of hard gates quickly identifies items that cannot be reliably forecast. Items failing these gates are immediately classified as "Under-Observed" with a capped score proportional to how close they are to the threshold. Gates check for insufficient observation periods, insufficient non-zero demand, high missing ratios, and zero total demand, each with frequency-specific thresholds.
6.3 Classification & Actionable Segmentation
| Score | Level | Segment | Recommended Action |
|---|---|---|---|
| 90–100 | Very High | Stable | Item-level forecasting with confidence |
| 75–89 | High | Stable | Standard forecasting methods |
| 55–74 | Medium | Volatile but Learnable | Forecast with wider prediction intervals |
| 30–54 | Low | Intermittent | Aggregate planning, wider safety margins |
| 0–29 | Very Low | Lumpy / Unforecastable | Safety-stock controls, do not forecast at this frequency |
6.4 AI-Powered Autonomous Analysis
The Forecastability Engine leverages an advanced AI Code Interpreter to execute the scoring methodology autonomously. SIA uploads the raw data, sends detailed analytical instructions, and the AI agent builds and runs the complete analysis pipeline: computing features, applying the scoring waterfall, generating per-item narratives, and exporting both a structured CSV ranking and a comprehensive markdown report.
Every item gets three narrative outputs: factual observations about its data characteristics, an interpretive rationale linking those observations to the forecastability verdict, and concrete next-step recommendations tailored to its specific classification.
Beyond Computation: Qualitative Reasoning at Scale
Beyond traditional ML, SIA integrates advanced AI reasoning and computer vision to add capabilities that were previously only available from human analysts.
7.1 Vision-Guided Forecasting
An optional Vision layer generates diagnostic chart views of the time series (raw observations, distributional profiles, and pattern decompositions) and submits them to a vision-capable AI model for structured analysis. The model produces chart-level observations (trend, seasonality, noise assessment) and synthesizes them into forecast adjustments: recommended model complexity, outlier sensitivity, confidence calibration, and damping parameters.
This creates a feedback loop where visual pattern recognition informs quantitative model configuration, bridging the gap between exploratory visual analysis and automated pipeline execution.
Visual Validation Charts
Actual vs. predicted overlays, residual distributions, and forecast fan charts are generated automatically and embedded directly in the HTML output.
Model Comparison Dashboard
Side-by-side metric tables showing ML, Physics, and Hybrid performance across MAPE, sMAPE, wMAPE, MAE, RMSE, MedianAE, and TrimmedMAPE.
7.2 Intelligent Schema & Frequency Inference
When users upload data without explicit column mappings, SIA's AI layer inspects headers and sample rows, returning structured analysis that identifies the date column, target variable, grouping columns, and the most appropriate analysis frequency. This eliminates a common friction point in onboarding new datasets, whether they come from Concinity exports, ERP extracts, or manual spreadsheets.
7.3 Automated Report Generation
The reporting pipeline produces richly formatted HTML reports that include validation charts, metric summaries, model comparison tables, and forecast visualizations. Reports are generated through a combination of template-driven rendering and AI-assisted narrative generation, ensuring that every output tells a coherent analytical story, not just numbers.
Forecasts That Reach Production Must Be Trustworthy
SIA includes multiple safety layers that prevent pathological outputs from propagating into downstream planning and execution systems.
Forecast Guard
Constrains period-over-period forecast changes to a configurable maximum ratio. A kill-switch triggers when validation error exceeds a safety threshold, falling back to simpler models automatically.
MAPE Rescue
A post-hoc calibration pass that applies global scale correction when the blended forecast's error exceeds a target threshold. Runs iterative passes until the target is met or no further improvement is possible.
Confidence Bounds
Applies statistical confidence intervals using configurable standard-deviation multipliers, ensuring that forecast uncertainty is always communicated alongside point predictions.
HFT-Inspired Volatility Monitoring
Computes realized volatility, higher moments (skew, kurtosis), and jump indicators across rolling windows. A regime detection system classifies low, normal, or high volatility for regime-conditional model weighting.
Deterministic Reproducibility
All AI calls support response caching with content-addressed hashing, configurable parameters, and global random seed control, ensuring identical inputs produce identical outputs across runs.
Closing the Loop Between Prediction and Action
Addverb's operational data, combined with SIA's forecasting intelligence, creates an end-to-end pipeline from demand signal to warehouse action. Concinity, Addverb's Warehouse Execution System, is the bridge that makes this possible. This partnership focuses specifically on SIA's forecasting and forecastability capabilities, where Addverb's domain expertise and data infrastructure are most directly relevant.
Where Addverb's Data Powers SIA
Addverb's deployments (ASRS, AMRs, conveyors) are where the demand signal lives. Concinity holds the task history, inventory movement records, inbound/outbound order flows, and shift-level throughput data that SIA needs to train and run forecasts.
Concinity already integrates upstream with ERP systems (SAP EWM and others) and downstream with automation hardware across multiple sites. This existing connectivity makes Concinity the natural bridge between SIA's forecast outputs and the live operations where those forecasts drive action.
Addverb understands shift structures, structural zeros (facilities closed on certain days or hours), equipment downtime patterns, and seasonal throughput cycles. All of these are the structural rules that SIA's physics-informed model depends on.
With live Concinity deployments across pharma, FMCG, e-commerce, and petroleum, Addverb offers SIA a ready pipeline of real-world pilots and labelled warehouse demand data: an opportunity to validate and refine SIA's models across diverse industry contexts.
Intelligence-aided motion: Addverb's operational data combined with SIA's forecasting intelligence has immense potential. When every robot, conveyor, and shuttle in a warehouse is informed by a demand forecast that understands the facility's operational rhythms, the result is motion guided by intelligence rather than reaction.
How SIA Creates Value Within Concinity
| Concinity Capability | How SIA Plugs In | Value Created |
|---|---|---|
| Inventory & Slotting | SIA demand forecasts feed directly into slotting decisions | High-velocity SKUs pre-positioned in accessible locations before peaks |
| Task Scheduling | Shift-level forecasts inform how tasks are queued and prioritized | Better labour and equipment utilization per shift |
| Replenishment Triggers | Forecast output drives when replenishment orders are raised | Fewer stockouts, less safety stock held |
| Upstream ERP Integration | Concinity's existing ERP connectors carry SIA outputs back to SAP/EWM | Closed-loop between forecast and procurement |
Orchestration Over Algorithms
SIA's effectiveness comes not from any single technique, but from the careful orchestration of complementary methods with built-in quality assurance.
Hybrid > Single Model
By blending ML and Physics tracks, SIA captures both data-driven patterns and domain-informed structure, consistently outperforming either approach in isolation.
Assess Before You Forecast
The Forecastability Engine prevents wasted effort by quantifying which items have sufficient data quality and signal to justify automated forecasting.
Frequency-Aware Throughout
From outlier thresholds to minimum history requirements to disaggregation profiles, every component adapts its behavior to the data's native frequency.
AI Reasoning, Not Just Computation
The AI Vision layer and autonomous analysis add a qualitative reasoning dimension that catches patterns and anomalies that purely quantitative methods miss.
Comprehensive Metric Coverage: SIA evaluates forecasts across seven complementary metrics (MAPE, sMAPE, wMAPE, MAE, RMSE, MedianAE, and TrimmedMAPE) because no single metric tells the complete story. The priority metric is automatically selected based on data characteristics: wMAPE when zero-heavy, MAPE for clean series, and sMAPE as a robust default.
From Data to Decisions in Three Steps
No data science team required. SIA is designed so operations managers, supply chain leaders, and business analysts can generate production-quality forecasts without writing a single line of code.
Upload Your Data
Drop your historical data file into SIA. It accepts a CSV export, an Excel download from your ERP, or a Concinity data extract. The platform automatically detects columns, dates, frequency, and data structure.
Ask Your Question
Tell SIA what you need in plain language ("forecast daily demand for the next 90 days" or "which SKUs are forecastable?") and the platform configures itself accordingly.
Act on the Results
Receive interactive reports with clear forecasts, confidence intervals, and actionable segmentation. Export results into Concinity, your ERP, or any planning workflow.
Whether you're managing warehouse throughput, planning SKU replenishment, or optimizing shift-level resource allocation, SIA's forecasting and forecastability capabilities deliver the analytical depth of a custom-built data science solution with the simplicity your operations team needs. Built by ResEt AI and validated with Addverb's real-world logistics operations, these capabilities are ready to drive measurable improvement from day one.