Turn your business data into decisions — no technical expertise required.
ClarityAI is an AI-powered forecasting and analytics platform built for small business owners. Upload a CSV or Excel file, answer three guided questions, and get enterprise-grade forecasts, anomaly alerts, and plain-English insights in under five minutes.
Small business owners are stuck between two worlds that don't work for them.
Too simple: Excel and Google Sheets can visualize past data but have no predictive power. They show you what happened — not what's coming.
Too complex: Power BI and Tableau require dedicated data teams, weeks of setup, and technical expertise most small businesses don't have. Tableau starts at $70/user/month and still won't tell you what your data means.
The gap: There is no accessible tool that takes a non-technical business owner from raw data to a statistically valid forecast, anomaly detection, and plain-English recommendations — in minutes, not weeks.
ClarityAI fills that gap.
No configuration knowledge required. ClarityAI auto-detects your date column, metric column, and grouping from your file. You confirm in three clicks and choose how far ahead to forecast.
ClarityAI runs seven forecasting models simultaneously and combines them using a trained meta-learner that learns the optimal weight for each model on your specific dataset. Every forecast includes confidence bands and a reported accuracy score (SMAPE).
Three independent anomaly detectors (Isolation Forest, One-Class SVM, and Robust Z-score) run on the residuals of your time series. A data point is only flagged as an anomaly when at least two methods agree — minimising false alarms. Anomalies appear as markers on the forecast chart with a tooltip showing the date, actual value, expected value, and Z-score.
Beyond one-off anomalies, ClarityAI detects permanent structural shifts in your data — moments where the underlying trend changed (a new competitor, a price change, an external shock). These appear as shaded bands on the chart, distinct from anomaly markers.
ClarityAI generates 4–6 insight bullets from your forecast and anomaly results. Examples:
- "Upward trend of +18% forecast over the next 6 months"
- "Peak expected in April at 4,200 units — 22% above current"
- "2 anomalies detected — August 2023 was 2.3σ below expected"
- "Structural shift detected in March 2022 — mean increased by 31%"
- "Best model: Prophet with 91.4% accuracy (SMAPE)"
Four auto-computed KPIs update live above the chart: forecasted growth %, peak predicted month, ensemble accuracy score, and total projected value.
If your dataset has a group column (product, region, store), ClarityAI runs a separate forecast per group and lets you switch between them with a dropdown. A summary table ranks all groups by forecasted growth.
Download a PDF report containing the forecast chart, KPI cards, anomaly list, insight bullets, and model accuracy comparison — ready to share with stakeholders.
| Model | Type | Why it's included |
|---|---|---|
| Prophet | Additive decomposition | Handles holidays, trend shifts, and missing data robustly |
| SARIMA | Classical seasonal | Strong baseline for regular seasonal patterns |
| LightGBM | Gradient boosting | Captures complex non-linear trends via lag features |
| XGBoost | Gradient boosting | Diverse boosting implementation — improves ensemble quality |
| Holt-Winters (ETS) | Exponential smoothing | Fast, reliable baseline for multiplicative seasonality |
| Theta | Linear decomposition | Won the M3 International Forecasting Competition |
| N-BEATS | Deep learning | Pure neural approach — no feature engineering required |
| Adaptive meta-learner | Backtested ensemble | Learns optimal model weights per dataset via backtesting |
The adaptive meta-learner is the core academic contribution of this project. Rather than averaging model outputs equally, it holds out a recent window, scores every model on it, drops models far weaker than the best, and combines the survivors by weights inversely proportional to their backtested error — discovering the combination that minimises error for each specific dataset.
We evaluated several combiners on the M4 monthly benchmark (30 series, 18-month horizon). A Ridge stacking regression — the conventional stacked-generalisation choice — proved to overfit the single short holdout window and underperformed the simpler robust combiner. The inverse-error weighting plus weak-model pre-filter reduced mean SMAPE from 14.86% (Ridge) to 14.20%, beating every individual model including Theta (the M3 winner). The combiner is configurable via the ENSEMBLE_WEIGHTING environment variable (inverse_smape default, ridge, or shrink).
| Method | Approach | Catches |
|---|---|---|
| Isolation Forest | Tree-based outlier isolation | Global outliers in residual space |
| One-Class SVM | Boundary-based novelty detection | Distributional anomalies |
| Robust Z-score (MAD) | Statistical threshold (3.5σ) | Extreme residual spikes |
A point is flagged only when at least 2 of 3 methods agree — this consensus approach dramatically reduces false positives compared to any single method.
| Layer | Technology |
|---|---|
| Backend API | FastAPI + Uvicorn |
| ML pipeline | Python — Prophet, statsmodels, LightGBM, XGBoost, scikit-learn, neuralforecast |
| Anomaly detection | scikit-learn, scipy |
| Changepoint detection | ruptures (PELT algorithm) |
| Frontend | React 18 + Vite |
| Charts | Recharts |
| HTTP client | Axios |
Thesis notes, methodology, and evaluation reports are in docs/ (root README.md stays here).
This project was developed as a graduation capstone. The core research contribution is the adaptive meta-learner ensemble — a stacked generalisation approach that learns per-dataset model weights through backtesting, outperforming naive equal-weight averaging across diverse time-series datasets. The anomaly detection pipeline extends the consensus approach described in the literature: combining tree-based, kernel-based, and statistical methods reduces false positive rates compared to any single detector, particularly on short monthly business series where individual methods are prone to over-flagging.