███  █████                ████ 
         ░░░  ░░███                ░░███ 
  █████  ████  ░███████  █████ ████ ░███ 
 ███░░  ░░███  ░███░░███░░███ ░███  ░███ 
░░█████  ░███  ░███ ░███ ░███ ░███  ░███ 
 ░░░░███ ░███  ░███ ░███ ░███ ░███  ░███ 
 ██████  █████ ████████  ░░███████  █████               meow!
░░░░░░  ░░░░░ ░░░░░░░░    ░░░░░███ ░░░░░           ╱|、
                          ███ ░███               (˚ˎ 。7  
                         ░░██████                 |、˜〵 
                          ░░░░░░                  じしˍ,)ノ 

Tell us what the future holds, so we may know that you are gods.

Notebook-friendly, dataframe-oriented forecasting for Haskell

Sibyl is designed first for interactive statistical exploration. Import one module, fit a model, predict a horizon, and inspect the result:

import Sibyl

model = fit defaultNaiveSettings sampleTimeSeries
forecast = predict 12 model

The notebook facade deliberately reads like R or Python while the underlying direct modules retain checked Either results for applications. Sibyl is narrower than a general statistics toolkit: the goal is a coherent R-forecast-style workflow that starts and ends with Haskell DataFrames.

The current 0.0.0.1 surface is a stabilized foundation, not the full vision. Naive forecasting is the only release-ready model family today. ARIMA, Holt-Winters, decomposition, plotting, and dataframe-native grouped workflows remain under development and are not exposed as working features.

What works now

• Opaque TimeSeries index value values with checked construction (non-empty, equal-length vectors and a strictly increasing index).
• Safe slicing, lag/lead, ordinary and seasonal differencing, and rolling aggregations.
• Rolling mean, variance, standard deviation, sum, minimum, maximum, median, and Pearson correlation.
• Simple moving average and manually or automatically tuned single exponential smoothing.
• Last-value, mean, drift, and seasonal naive forecasts with prediction intervals.
• Fitted values, residuals, model summaries, and MAE/RMSE/MAPE/MASE.
• Basic conversion of one index column and one numeric value column to/from DataFrame.
• Generic notebook operations like fit, predict, predictWith, summarize, fitted, and residuals

Quick start

Notebook and REPL use

import Sibyl is the primary exploration interface.

import qualified Data.Vector.Unboxed as U
import Sibyl

sales = mkTimeSeries
  (U.fromList [1..8 :: Int])
  (U.fromList [10, 20, 30, 40, 12, 22, 32, 42 :: Double])

settings = defaultNaiveSettings
  { naiveMethod = Seasonal
  , period = Just 4
  , naiveCiLevel = 0.95
  }

model = fit settings sales
forecast = predict 6 model

forecastValues = observations (predPoint forecast)
lowerBounds = observations (predLower forecast)
upperBounds = observations (predUpper forecast)

-- In IO / an IHaskell cell:
-- summarize model

The central vocabulary is model-independent:

fit         settings series
predict     horizon model
predictWith horizon futureData model
summarize   model
fitted      model
residuals   model

fit infers the model family from the settings value. predict is for models that need no future inputs. predictWith supports future regressors or other model-specific future data without complicating the common case.

Model-specific helpers such as fitNaive and forecastNaive remain as optional conveniences, if you prefer them.

Safe use

Import direct modules when you need explicit error handling. They expose the same implementation through Either:

import Sibyl.Model (Prediction(..))
import Sibyl.Models.Naive (fitNaive, predictNaive)
import Sibyl.TimeSeries (index, observations, sampleTimeSeries)

main :: IO ()
main = case fitNaive sampleTimeSeries >>= predictNaive 3 of
  Left err -> print err
  Right prediction -> do
    print (index (predPoint prediction))
    print (observations (predPoint prediction))

DataFrame status

Basic interop exists today:

fromDataFrame :: Text -> Text -> DataFrame -> Either ConversionError (TimeSeries index Double)
toDataFrame   :: TimeSeries index value -> DataFrame

but this is first priority for the future.

Modules

• Sibyl is the primary notebook/REPL facade with generic fit and predict operations.
• Sibyl.TimeSeries has checked series construction and transformations.
• Sibyl.Smoothing is moving-average and single-exponential smoothing.
• Sibyl.Accuracy has forecasting accuracy measures.
• Sibyl.Model contains the common fitted-model, prediction, summary, and error types.
• Sibyl.Models.Naive has the "safe" naive model fitting and prediction.

Near-term roadmap

1. DataFrame 1.1 migration and first-class single/grouped frame workflows.
2. Forecast frequency/calendar policies that replace implicit Enum index extension.
3. ACF/PACF, Ljung-Box, and reference-value diagnostic tests.
4. Optimized Holt/Holt-Winters with intervals.
5. ARIMA/SARIMA/SARIMAX and deterministic automatic order selection.
6. Rolling-origin cross-validation and accuracy by horizon.

Building

Sibyl supports GHC 9.6.7 and 9.10.3.

cabal build all --enable-tests
cabal test all

License

BSD-3-Clause. See LICENSE.