GRIM Forecast Studio
Inventory Demand Forecasting — Train, Test, Deploy, Monitor
Total Series
-
SKU-Location combinations
Forecast Horizon
24w
Weekly predictions ahead
Algorithms
12
Competing models
Last Selection
-
Model selection cycle
Predictions
-
Total forecast rows

Forecast Overview

Series Classification Distribution

Model Assignment Distribution

Pipeline Architecture

1
Ingest
DuckDB
2
Build
Time Series
3
Classify
4-Tier
4
Select
MAPE Arena
5
Forecast
8 Algorithms
6
Output
Post-Process
7
Monitor
Error Track

Series Explorer

Run a forecast first

Click "Run Forecast" to generate predictions

📈

Select a Series

Click a series from the list to view its forecast

Model Arena — Algorithm Competition

Head-to-Head Win Rate Matrix

Series Classification

SeriesClassZero%WeeksModel

Forecast Configuration

Forecast Parameters

Horizon24 weeks
Selection Frequency30 days
CPU Cores23

Classification Thresholds

Dead Lookback180d
Intermittent Threshold0.6
Dead Threshold0.95

LightGBM Hyperparameters

Estimators300
Learning Rate0.05
StrategyRecursive

Algorithm Registry

LightGBMON
XGBoostON
AutoARIMAON
AutoETSON
AutoCESON
AutoThetaON
NaiveON
IMAPAON
CrostonON
SBAON
TSBON
QuantileON

Error Monitoring & Degradation

Degradation Threshold
20%
MAPE increase triggers flag
Lookback Window
4w
Recent evaluation window
Log Retention
12w
Prediction history kept

Monitoring Pipeline

1. Log Predictions
Save each cycle's forecasts with dates
2. Compute Recent MAPE
Match predictions to actuals over last 4 weeks
3. Flag Degraded Series
Compare to baseline — flag if MAPE grew >20%
4. Priority Re-Selection
Flagged series get priority in next model selection

Train & Deploy

Quick Forecast Run

Run forecasts using cached model assignments (fast). Use when model selection is up to date.

Full Re-Train & Deploy

Force full model selection competition (MAPE arena), then forecast. Slower but ensures best models are selected.

Pipeline Run Log

Ready to run. Click a button above to start the forecast pipeline.

Algorithm Feature Summary

AlgorithmTypeSeries ClassSpeedStrengths
LightGBMML / TreeACTIVEFastGeneral demand, sparse data, interpretable
AutoARIMAStatisticalACTIVEMediumStationary, mildly seasonal series
AutoETSState-spaceACTIVEMediumClear trend + seasonality
AutoCESComplex Exp.ACTIVEMediumComplex seasonal patterns
AutoThetaDecompositionACTIVEFastM3/M4 competition winner
NaiveBaselineALLInstantUniversal fallback, no-pattern safety net
IMAPAIntermittentINTERMITTENTMediumLumpy/sparse demand (>60% zeros)
QuantileML / QuantileACTIVEFastP10/P50/P90 prediction intervals