Statistical Forecasting Models Overview
Our statistical forecasting engine is designed to operate on time-series data—sequences of observations represented simply as (date, value) pairs. Each forecasting model ingests these historical data points and produces future data points in the same structure.
We currently support six industry-standard forecasting models, each suited to different demand patterns and business requirements. By default, Confido automatically optimizes parameters.
1. Moving Average
The Moving Average model produces forecasts by averaging a fixed number of the most recent observations. It is the simplest of our supported models.
Parameters
window_size — Number of past weeks included in the moving average.
2. Single Exponential Smoothing (SES)
Single Exponential Smoothing is one of the most widely used techniques in statistical forecasting. This is effectively a smart moving average that reacts to data changes.
SES applies a weighted average where more recent observations are given greater importance. Its core formula is:
This means each new prediction is a blend of the most recent observed value and the previous prediction.
SES produces a constant-level forecast.
Parameters
alpha — Smoothing factor (0–1).
SES is technically a special case of the Holt–Winters method with no trend or seasonality.
3. Holt–Winters (Triple Exponential Smoothing)
Holt–Winters is an extension of SES that additionally models trend and seasonality, making it suitable for more complex demand patterns.
Parameters
alpha — Level smoothing factor
beta — Trend smoothing factor
Not applicable when
trend_type = none
gamma — Seasonal smoothing factor
Not applicable when
seasonality_type = none
trend_type — One of:
auto(Confido picks the best)additive,multiplicative,noneseasonality_type — One of:
auto(Confido picks the best)additive,multiplicative,none
These settings determine how strongly the model responds to trend and seasonal variations.
4. Croston’s Method
Croston’s method is specifically designed for intermittent demand—situations where non-zero observations occur infrequently (e.g., sporadic orders).
Note: Markov Chain is also great for intermittent demand
It is a variant of SES that separately smooths the size and occurrence interval of non-zero events.
Croston’s method produces a constant-level forecast.
Parameters
alpha — Smoothing factor
5. SARIMA (Seasonal ARIMA)
Seasonal ARIMA (SARIMA) is a powerful and flexible statistical model used for time-series forecasting. It is capable of modeling recurring patterns—such as monthly demand spikes or yearly cycles—along with trends and autocorrelation in the data.
SARIMA is widely used in academic and statistical fields and is considered a gold standard for traditional time-series analysis. While it is less commonly used day-to-day within CPG organizations relative to simpler smoothing-based models, it provides strong interpretability and is especially useful for analysts who require explicit control over model structure.
What SARIMA Captures
SARIMA can model four major components of a time series:
Level — the baseline value over time
Trend — long-term upward or downward movement
Seasonality — recurring patterns (weekly, monthly, yearly, etc.)
Autocorrelation — the degree to which past values influence future values
This makes SARIMA appropriate for time series where:
Seasonal patterns repeat predictably
Trends evolve gradually
Noise is correlated (i.e., demand today depends on demand last week)
How SARIMA Works
A SARIMA model is defined by a set of non-seasonal and seasonal parameters:
Non-seasonal parameters:
p — number of autoregressive lags
d — degree of differencing to remove trend
q — number of lagged forecast errors (moving average terms)
Seasonal parameters:
P — seasonal autoregressive order
D — seasonal differencing order
Q — seasonal moving average order
m — seasonality period (e.g., 12 for monthly data with yearly seasonality)
These parameters define the structure of the model and allow for highly customizable forecasting behavior.
These parameters are typically selected via statistical diagnostics (e.g., ACF/PACF plots) or automated selection methods ("auto-SARIMA").
6. Markov Chain
Our Markov Chain model forecasts demand by learning state-to-state transition probabilities. For example, given that an account shipped 10 units last month, the model computes the likelihood of shipping 20 units next month based on historical transitions.
Markov chain cannot capture trend, but can capture seasonality and intermittent demand.
This produces forecasts that visually mimic historical patterns and variability.
Parameters
None — The model behavior is deterministic given the input series.