End-to-end ML app for predicting annual insurance charges with SHAP explainability and LLM-powered interpretation for business and technical users.

InsureCast
Time-series forecasting dashboard for monthly insurance claims and costs, with scenario stress-testing and AI-generated summaries
Highlights
Teams planning reserves and operations need consistent forward views of claim frequency and cost, but spreadsheets and one-off models are hard to reproduce, segment, and stress-test.
Build a forecasting prototype that predicts monthly claims and costs with uncertainty bands, supports what-if scenarios, and ships an aligned natural-language summary in an interactive UI.
Built per-segment SARIMAX seasonal models with ARIMA/seasonal-naive fallbacks, a scenario API for severity and frequency shocks, and OpenAI summaries tied to the same rows the charts use.
Replaced ad-hoc spreadsheets with one API and a single definition of “forecast”: refresh by segment and horizon, apply shocks, and read a chart-aligned narrative, all in one dashboard.
Core Team
Overview
InsureCast is a time-series forecasting dashboard that projects monthly insurance claims and paid amounts, giving teams that plan reserves and operations a consistent, reproducible forward view rather than static history. Users pick a segment and horizon, and the system returns point forecasts with confidence-style bands, a monthly data table, and a short narrative aligned to the very same numbers shown on the charts.
Built as a FastAPI backend and Next.js frontend over shared demo data, it fits per-segment SARIMAX seasonal models with automatic fallbacks, and exposes a scenario API for stress-testing. A what-if panel lets analysts apply severity-inflation and frequency-shock adjustments and immediately see the forecast charts and table update, turning ad-hoc spreadsheet work into a single, API-first definition of "forecast."
Data
The demo is built on public OSHA Severe Injury Report (SIR) data, normalized into monthly aggregates the dashboard can serve directly:
- •OSHA SIR (demo): Public severe-injury reports ingested and normalized to the dashboard schema via dedicated ingest scripts.
- •Merged demo table: A single CSV of all segments × months (month, state, industry, claim type, counts, and baseline average cost), loaded by the API at startup.
- •Synthetic fill: An optional step completes coverage for demo segments when a series has gaps.
Segments are defined by state, industry, and claim type, each modeled as its own monthly time series.
Methods
The pipeline fits a seasonal model per segment and serves forecasts, scenarios, and summaries from the same underlying rows:
- Seasonal Modeling: Each segment's monthly series is fit with a SARIMAX (1,1,1)×(1,1,1,12) model, capturing trend and 12-month seasonality.
- Robust Fallbacks: When a series is too short or noisy for SARIMAX to fit, the system automatically falls back to ARIMA or a seasonal-naive method.
- Forecasts with Uncertainty: Point forecasts are produced with 95% prediction intervals and a clear historical-vs-forecast split for both claims and average cost per claim.
- Scenario Recalculation: A scenario endpoint applies severity-inflation and frequency-shock adjustments, refreshing the forecast segment on demand.
- Aligned AI Summary: An OpenAI-generated narrative (with a deterministic fallback when no key is configured) describes the selected window using the exact rows the charts consume.
Results
By unifying forecasting, scenarios, and reporting behind one API, InsureCast streamlines the planning loop:
- •Planning Cycle: Refresh views by segment and horizon in one UI instead of manually rebuilding them per question.
- •Consistency: One API and one merged definition of "forecast" replace scattered spreadsheets and divergent assumptions.
- •Scenarios: Severity and frequency shocks update the charts and table immediately, making side-by-side comparison easy.
- •Communication: Chart-aligned summary text replaces static slides, keeping the narrative tied to the numbers.
A walkthrough of selecting a segment, generating monthly claims and cost forecasts with prediction bands, and stress-testing the outlook under severity and frequency shocks is shown in Figure 1, and the full dashboard view is shown in Figure 2.
Conclusion
InsureCast shows how classical time-series modeling, an interactive scenario loop, and an aligned AI narrative can come together into a single, reproducible forecasting tool. Its time-series-first design (monthly frequency and seasonality with robust fallbacks) and API-first architecture make claims and cost projections consistent, segmentable, and easy to stress-test.
As a prototype, it is positioned to grow toward production: roadmap items include live production data feeds and authentication, per-tenant configurable models and horizons, and monitoring with CI hardening.
Keep Exploring
Non-invasive ICP monitoring using deep learning on cerebral blood flow signals from near-infrared photonic sensors.



