DOCSEARCHbyAlgolia
Submit Docs

AInventory Documentation

website·Indexed 31 days ago

https://docs-ainventory.directrixanalytics.com

Documentation for AInventory, an AI-powered inventory and demand planning platform. Includes guides on collaborative forecasting, SKU optimization, safety-stock policies, and the use of SofIA, the built-in AI copilot. Covers key metrics such as Forecast Value Added (FVA), Net Flow (FN), and the VN2 Inventory Planning Challenge.

Tokens
53.1K
Snippets
109
Records
431
Agent score
51%

What's inside AInventory

  1. Overview of AInventory Statistical Forecasting Models

    AInventory employs a statistical cascade of models to generate demand predictions:

    • ARIMA: Statistical time-series model capturing trend and autocorrelation.
    • CNNQR (Convolutional Neural Network — Quantile Regression): Deep-learning model estimating the full distribution of future demand (First level of cascade).
    • DeepAR: RNN-based probabilistic forecasting that learns global patterns across multiple SKUs (First level of cascade).
    • ETS (Error, Trend, Seasonality): Exponential smoothing model using decreasing weights on historical data (Second level of cascade).
    • Prophet: Meta-developed model handling trend, multiple seasonality, and calendar effects (Second level of cascade).
    • NPTS (Non-Parametric Time Series): Approach using empirical history directly without assuming a parametric distribution.
    • Rule 50/50: A fallback heuristic using a 50% weighted average of historical mean and most recent demand when other models are unreliable.

  2. Overview of SofIA AI copilot in AInventory

    SofIA is an AI copilot built into AInventory designed to help users validate adjustments, interpret KPIs, analyze replenishment, and query data using natural language. It is available across the Forecast, KPIs, Inventory, FVA, and History modules. Responses are streamed in real time. Conversations reset automatically upon logout or after 2 hours of inactivity.

  3. Overview of SofIA AI Copilot in AInventory

    SofIA is an integrated AI copilot in AInventory used to validate settings, interpret KPIs, analyze replenishment, and ask natural language questions about data within the active module. It is available in the Forecast, KPIs, Inventory, FVA, and History modules. Responses are streamed in real-time. Conversations are automatically reset upon logout or after 2 hours of inactivity.

  4. Understand Safety Stock concepts and basic formulas

    Safety stock (SS) is the inventory maintained above the expected demand during the lead time and review period to act as a buffer against uncertainty.

    For policies using a factor k, the relationship is defined as: SS = k · σ_W

    Where:

    • σ_W: The variability of demand in window W (Lead Time + Review Period).
    • k: A multiplier where a higher value increases safety stock and fill rate, while a lower value reduces tied-up inventory but increases stockout risk.
    SS ≈ target stock B − expected demand in window W

  5. Understand the Four-Level Statistical Cascade for SKU forecasting

    The system automatically assigns a statistical method to each SKU based on the number of observations (n) in its historical error matrix. This 'cascade' ensures that the complexity of the statistical model matches the available data volume, moving from high-assumption models (with little data) to data-driven non-parametric models (with abundant data). The thresholds between these levels are configurable per instance.

  7. Understand AInventory's recommendation transparency and traceability

    AInventory is designed as an auditable assistant rather than a black box. Every recommendation is accompanied by explicit metadata to prevent blind trust in the numbers. For every processed SKU, the system exposes:

    1. The assigned statistical cascade level and the number of observations (n).
    2. The chosen distribution (Normal, Gamma, Lognormal, or non-parametric bootstrap) and the selection criterion used (AIC, Shapiro-Wilk, or default).
    3. The winning policy and optimal k*, including the expected cost ranking of all five evaluated policies.
    4. Active warnings with their specific codes and severity levels.
    5. Forecast parameters and historical errors to identify potential bias.

  8. Distinguish between Accuracy and Training metrics in forecasting

    Forecasting evaluation is split into two distinct families to avoid incorrect conclusions about model performance versus human intervention:

    1. Accuracy metrics (BIAS, MAE, MAPE, RMSE, Accuracy): Evaluate the final forecast delivered to the business, including all manual adjustments. These measure the success of the entire S&OP process.
    2. Training metrics (MASE, WAPE, wQL, MSE): Evaluate only the quality of the underlying statistical model before human adjustments. These indicate if the algorithm learned correctly from historical data and help in comparing candidate models or detecting overfitting.

  9. Understand SofIA Quick Actions general behavior

    Quick actions are contextual suggestions displayed automatically by SofIA based on the active module. They allow users to generate analytical responses about current data without typing manual questions.

    Key system behaviors include:

    • Streaming Output: Responses are delivered progressively.
    • Formatting: SofIA supports Markdown (tables, lists, bold text).
    • Session Timeout: Sessions reset upon closing the application or after 2 hours of inactivity.

  10. Understand the four-level statistical cascade for SKU forecasting

    The system uses a four-level statistical cascade to assign the most appropriate forecasting method to an SKU based on the number of observations (n) in its historical error matrix. The core principle is that as more data becomes available, the system reduces distributional assumptions and migrates to more precise, non-parametric methods automatically.

  11. Understand the VN2 Inventory Planning Challenge validation

    The VN2 Inventory Planning Challenge (2025-2026) is an international competition used as external evidence to validate the 'predict-then-optimize' + simulation paradigm. In this competition, independent teams used real historical data to produce replenishment plans. The winning approach achieved a total cost reduction of approximately 13% compared to the reference method. Crucially, no winning team used the classic safety stock formula (z · σ_d · √L), instead opting for scenarios of future demand and policy evaluation based on expected total cost minimization—the same paradigm implemented by AInventory.