capacity-forecaster

A robust capacity planning forecasting engine for workforce and operations teams. Uses SARIMAX with automatic model selection to produce multi-month forecasts of Volume, Hours, FTE, and Shrinkage-adjusted FTE — with 95% confidence intervals — for any number of planning groups.

---

Features

• Automatic model selection — grid-searches SARIMAX orders and picks the best fit by AICc (corrected AIC), per group, per metric
• Shrinkage forecasting — when shrinkage history is supplied, it is modelled as its own time series so seasonal leave patterns flow through to adjusted headcount
• Gap filling — missing months within a group's history are detected and linearly interpolated automatically, with the count reported in Imputed_Months
• Data quality flagging — groups with limited history are forecasted and flagged LOW_HISTORY rather than silently dropped, so you decide what to trust
• Confidence intervals — every metric includes _Lower and _Upper bounds at the 95% level
• Flexible input — column names are ignored; columns are matched by position, so your existing DataFrames work without renaming

---

Installation

pip install capacity-forecaster

---

Quickstart

import pandas as pd
from capacity_forecaster import CapacityForecaster

df = pd.read_csv("capacity_planning_data.csv", parse_dates=["date"])

forecaster = CapacityForecaster(
    weekly_hours=37.5,       # contracted hours per FTE per week
    forecast_horizon=12,     # months ahead to forecast
    default_shrinkage=0.30,  # fallback shrinkage rate (30%)
)

results = forecaster.forecast(df)
print(results.head())

---

Input format

Columns are matched by position, not by name. Pass your DataFrame with columns in this order:

Position	Role	Type	Required	Description
0	Date	datetime	✅	Any monthly date (month-start or month-end). Mixed formats within a column are fine.
1	Group	str	✅	Group identifier (e.g. team or queue name).
2	Volume	float	✅	Units of work completed that month.
3	Hours	float	✅	Total workload hours that month.
4	Shrinkage	float	➕ optional	Shrinkage rate in [0.0, 1.0), e.g. 0.25 = 25%.

Each row is one group × one month. You need at least 6 months of history per group to attempt a forecast; 36 months is recommended for full seasonal accuracy.

Because columns are positional, you can pass a subset of your DataFrame directly:

# Your columns can be named anything
results = forecaster.forecast(
    df[["period_end", "team", "contacts", "handle_hours"]]
)

# With optional shrinkage
results = forecaster.forecast(
    df[["period_end", "team", "contacts", "handle_hours", "shrinkage_rate"]]
)

---

Output columns

Column	Description
Date	Forecast month (month-start).
(your group column name)	Group label, using whatever name your input column had.
Forecasted_Volume	Predicted work volume.
Forecasted_Volume_Lower	Lower 95% CI bound.
Forecasted_Volume_Upper	Upper 95% CI bound.
Forecasted_Hours	Predicted workload hours.
Forecasted_Hours_Lower	Lower 95% CI bound.
Forecasted_Hours_Upper	Upper 95% CI bound.
Forecasted_FTE	Raw FTE required (hours ÷ hours-per-FTE-per-month).
Forecasted_FTE_Lower	Lower 95% CI bound.
Forecasted_FTE_Upper	Upper 95% CI bound.
Forecasted_FTE_Adjusted	FTE grossed up for shrinkage (Raw FTE ÷ (1 − Shrinkage)).
Forecasted_FTE_Adjusted_Lower	Lower 95% CI bound.
Forecasted_FTE_Adjusted_Upper	Upper 95% CI bound.
Shrinkage_Used	The shrinkage value applied each forecast month (forecasted or default).
Imputed_Months	Number of months gap-filled within the group's history.
Data_Quality	"OK" (≥ 36 months history) or "LOW_HISTORY" (6–35 months).

---

Data quality

History	Behaviour	Data_Quality
< 6 months	Forecast attempted with simplified model. Results should be treated with caution.	LOW_HISTORY
6–35 months	Forecast attempted. Non-seasonal ARIMA candidates used where seasonal fit is unreliable.	LOW_HISTORY
≥ 36 months	Full seasonal SARIMAX with complete candidate grid.	OK

Filter to only reliable forecasts when needed:

reliable = results[results["Data_Quality"] == "OK"]

---

Parameters

CapacityForecaster

Parameter	Type	Default	Description
weekly_hours	float	37.5	Contracted hours per FTE per week.
forecast_horizon	int	12	Months ahead to forecast.
default_shrinkage	float	0.30	Fallback shrinkage when no shrinkage column is supplied or values are NaN.

forecast(df)

Runs the forecast pipeline and returns a pd.DataFrame with one row per group × forecast month.

---

Shrinkage

Shrinkage represents time lost to leave, training, breaks, and absence. The adjusted FTE formula is:

Adjusted FTE = Raw FTE / (1 - Shrinkage)

Without a shrinkage column — default_shrinkage is applied flat across all forecast months for all groups.

With a shrinkage column — shrinkage is forecasted independently per group using SARIMAX, capturing seasonal patterns (e.g. higher shrinkage in August and December due to holiday leave). The fallback chain if the model cannot fit is: historical group mean → default_shrinkage. The forecasted value is clipped to [0.0, 0.9999] to keep availability positive. The value actually applied each month is always visible in Shrinkage_Used.

---

Public API

from capacity_forecaster import (
    CapacityForecaster,       # main class
    MIN_DATA_POINTS,          # 36 — recommended minimum observations
    ABSOLUTE_MIN_DATA_POINTS, # 6  — hard floor for attempting a forecast
    CONFIDENCE_LEVEL,         # 0.95
    DataQuality,              # DataQuality.OK / DataQuality.LOW_HISTORY
)

---

Dependencies

• pandas >= 2.0
• numpy >= 1.23
• statsmodels >= 0.14

---

License

MIT