The rule-based statistical expert system behind RoadMap's software and consulting, and the forecasting engine Oracle has licensed since 1999. For every time series, Geneva tests 10 statistical methods, filters outliers, detects seasonality, tunes parameters, and selects the best fit.
Engine Architecture & Performance
A deterministic, rule-based statistical expert system engineered in C++. Every forecast is explainable, reproducible, and scored with holdout diagnostics.
Geneva inspects series length, zero-demand intervals, variance stability, and spectral frequency to classify the underlying demand structure before fitting.
Automated pre-fit filters detect seasonal cycles and isolate one-time promotional spikes or stockout dips so a single anomalous period never distorts the baseline.
Instead of applying static textbook defaults, an optimized C++ search tunes smoothing weights, damping factors, and regression coefficients specifically for the series.
Predefined expert decision rules built from decades of cross-client studies select the winning model and return MAPE, MAD, RMSE, Tracking Signal, and prediction intervals.
Whether planners run interactive diagnostics in RoadMap GPS or developers call the REST endpoint from Python or Claude MCP, the same Geneva C++ core computes the forecast, fit metrics, and prediction bounds.
Full transparency into the selected method, holdout sample error (MAD, MAPE, RMSE, Tracking Signal), and 90% / 95% / 99% confidence intervals.
Pass a time series via REST, Python SDK, or MCP server and receive the best-fit forecast, fitted history, and prediction intervals in under 200 ms on average.
Six regression forms, three exponential smoothing variants, and Croston's method for intermittent demand. Model forms 1 to 9 match Oracle's OLAP DML documentation; Croston's method is available in the Geneva API and RoadMap suite.
6 Models · Trend & Lifecycle Shapes
Captures structural long-term trajectories including linear baselines, proportional power growth, compounding exponential growth or decay, logarithmic tapering, capacity ceilings, and S-curve product launches.
3 Models · Level, Trend & Seasonality
Weights recent observations dynamically across single smoothing, Holt’s linear trend, and Holt-Winters triple exponential smoothing with damped trend and multiplicative seasonality.
1 Model · Croston's Method
Decomposes sparse series into separate estimates of demand size and inter-arrival interval so slow-moving SKUs, spare parts, and specialty therapies are never over-forecasted.
| # | Method & Family | Model Form | How It Works & Best For |
|---|---|---|---|
| 01 | Linear regressionRegression | y = a·x + b | Fits a straight-line trend through time. The baseline every other method has to beat. Best for: Steady linear trend |
| 02 | Power regressionRegression | y = c·xᵃ | Captures growth whose rate scales with time, curving up or flattening depending on the exponent. Best for: Growth that accelerates or decelerates proportionally |
| 03 | Exponential regressionRegression | y = c·eᵃˣ | Models a constant percentage change per period: compounding growth, or steady decay for aging products. Best for: Compounding growth or decay (e.g. aging products) |
| 04 | Logarithmic regressionRegression | y = a·ln(x) + b | Rises quickly at first, then levels off. Typical of adoption that saturates gradually. Best for: Fast early growth that flattens |
| 05 | Asymptotic regressionRegression | y = x / (a + b·x) | Approaches a fixed upper limit, useful when capacity or market size caps demand. Best for: Growth approaching a ceiling |
| 06 | Exponential asymptotic (logistic / S-curve) regressionRegression | y = cKeᵃˣ / (1 + ceᵃˣ) | The S-curve: slow start, rapid adoption, then saturation. The classic shape of a product launch. Best for: Product launches and market saturation |
| 07 | Single exponential smoothingExponential smoothing | Weighted average of past values | Weights recent observations more heavily than older ones. Fast and robust for short horizons. Best for: Short-term, non-seasonal data |
| 08 | Double exponential smoothing (Holt's linear trend)Exponential smoothing | Smoothed level + trend | Adds a smoothed trend term to the level, so forecasts follow a series that is rising or falling. Best for: Trending, non-seasonal data |
| 09 | Holt-Winters (triple exponential smoothing)Exponential smoothing | Level + damped trend + multiplicative seasonality | Smooths level, a damped trend, and a multiplicative seasonal pattern together. The workhorse for seasonal demand. Best for: Seasonal data |
| 10 | Croston's methodIntermittent demand | Separate smoothing of demand size and interval | Forecasts how much and how often separately, for items that sell in irregular bursts with many zero periods. Best for: Intermittent / sparse demand (spare parts, slow movers) |
Born out of econometric research at Harvard Business School in the 1980s and battle-tested inside enterprise databases and Fortune 500 supply chains since 1988.
Founded 1988
At Harvard Business School in the 1980s, Rudolph Pizzano III, Ph.D. saw that the leading econometric models of the era, despite their theoretical complexity, consistently underperformed on messy real-world business forecasts. He built the statistical framework that became the Geneva Forecasting™ Engine to handle the thousands of edge cases classical models miss, founding RoadMap Technologies in 1988.
In the 2000s, Prof. Frank Wolak of Stanford University, a RoadMap board member, guided engine design through his research on electricity market deregulation, while Dr. Evdokia Nikolova (Ph.D., MIT) extended Geneva to high-frequency 15-minute interval electricity demand forecasts.
Partner Since 1995 · Licensed Since 1999
RoadMap's technology partnership with Oracle began in 1995. Oracle licensed Geneva in September 1999, shipped it in Oracle Express Server 6.3, and deployed it across the enterprise from 2000. Geneva served as the statistical core of Oracle OLAP, Oracle Demand Planning, and Oracle Retail Forecasting, and still ships today inside Oracle Database 26 OLAP DML.
“A best-of-breed statistical forecasting engine: Geneva Forecasting™ from Roadmap Technologies. Developed specifically for demand forecasting applications, this engine has multiple forecasting methods (linear and non-linear regression, variations of exponential smoothing), efficient parameter search and outlier detection algorithms, and a rule based system for automatic best fit forecasts based on time and frequency domain analysis and predefined decision rules generated from cross-client studies.”
Official Oracle documentation citations:
Access the Geneva Forecasting™ Engine through our developer API, browser workspace, enterprise Windows desktop platforms, or inside Oracle Database.
Embed sub-200 ms statistical forecasting directly into applications, data pipelines, and AI agents like Claude.
Benchmark the Geneva Expert System alongside 40+ statistical, ML, and deep learning models in the browser.
Enterprise demand planning software combining the ease of Excel, the statistical power of R, and SQL Server scale.
Full-lifecycle planning platform uniting Geneva best-fit forecasting, inventory tracking, and 360° executive visibility.
Licensed by Oracle since 1999 as the statistical forecasting engine inside Oracle Database OLAP DML.
Weekly NFL Moneyline, Spread, and Over/Under predictions powered by RoadMap’s statistical modeling framework.
Yes. Oracle licensed the Geneva Forecasting™ Engine in 1999. It was the statistical core of Oracle OLAP, Oracle Demand Planning and Oracle Retail Forecasting, and it still ships inside Oracle Database 26 OLAP DML.
Ten methods. Six regression forms (linear, power, exponential, logarithmic, asymptotic and logistic S-curve), single and double exponential smoothing, Holt-Winters, and Croston's method, plus automatic best-fit selection.
Croston's method forecasts intermittent demand, meaning items that sell in irregular bursts with many zero periods. It smooths demand size and the interval between demands separately, which suits spare parts and slow movers.
No. Geneva is a highly optimized C++ engine that runs on standard CPUs, averaging under 200 ms per forecast via the API.
Get a free API key with 1,000 forecasts a month, launch TrailBlazer in your browser, or talk to us about putting Geneva to work in your S&OP cycle.
