Skip to main content
CARLA HQ
DATASET PROFILE / TIME SERIES

Monash High-Density Time Series Benchmark

High-density multi-domain operational benchmark tracking 30 continuous daily time series over a two-year observation window. The benchmark targets a 28-day forecasting horizon to evaluate multi-step operational planning and zero-shot foundation model generalization.

Retail & Supply Chain Commerce & Industry use casesNACE G47Time Series ForecastingDaily Cadence28 Steps HorizonMarkdown ↗Target: targetfev / AutoGluon ↗
✨ Try in CarlaInstall Chrome Extension ↗100% In-Browser WebGPU • Zero Cloud Upload
21,900Records (Rows)
30Series Sequences
DailySampling Cadence
28 DaysForecast Horizon (10 Windows)
7d Weekly
Dominant Seasonality?7-day weekly operational cycle with distinctive weekday vs. weekend throughput variations and multi-month baseline trends across 730 historical steps.
41 Methodsfev Leaderboard (SQL)
BUSINESS CONTEXT & OBJECTIVESupply Chain Operations & Inventory Management (G47)

Business Objective: Monash High-Density Time Series Benchmark

01Business Context

Enterprise replenishment planning and operational demand scheduling operate on rolling 4-week (28-day) budgeting cycles across multiple fulfillment nodes. Planners need dependable daily projections to balance cross-dock allocations and warehouse staffing.

02Analytical Objective

Accurately forecast daily target volumes across 30 operational entities over a 28-day horizon (H=28) capturing dominant weekly day-of-week demand shifts.

03Economic & Decision Impact

Under-forecasting leads to immediate stockouts, missed delivery SLAs, and expedited freight surcharges, while over-forecasting incurs holding costs, working capital lockup, and warehouse capacity constraints.

04Temporal Dynamics & Covariate Information Flow

Dominant Seasonality: 7-day weekly operational cycle with distinctive weekday vs. weekend throughput variations and multi-month baseline trends across 730 historical steps.

Covariates & Information Flow: Governed by calendar attributes (day of week, month, payday cycles) and calendar holidays. Series telemetry captures continuous operational volume signals without auxiliary sensor covariates.

ML Benchmark Narrative

Classical baselines like Seasonal Naive and AutoARIMA capture the persistent 7-day cyclicality but struggle with subtle multi-month level shifts across the 28-day horizon. LightGBM with recursive lag features delivers competitive point-forecast accuracy, while zero-shot foundation models such as Timestral, TimesFM, and MOIRAI excel at zero-training in-context extrapolation without requiring hyperparameter tuning across disparate series scales.

Source Origin:Monash Time Series Forecasting Repository
Creator:Rakshitha Godahewa, Christoph Bergmeir, Geoffrey I. Webb, Rob J. Hyndman, and Pablo Montero-Manso (2021)
License:Creative Commons Attribution 4.0 International (CC-BY 4.0)
IN-BROWSER FORECAST DEMO

Forecast this dataset in your browser

A representative series is read directly from this dataset’s Parquet file. Move the forecast start, compare held-out observations, and toggle available numeric covariates with Timestral 5M (12 MB) or switch to ci-proto (~346 MB, WebGPU). Benchmark scores above use the stated evaluation setup.

Live sample from dataset Parquet
Timestral 5M · WebGPU
Nord Pool Day-Ahead Electricity PricesHourly · €/MWh
Forecast History
Loading sample data…
Covariates
FEATURE SPECIFICATION & DATA DICTIONARY

Target & Feature Column Definitions

Exhaustive business definitions, measurement units, roles, and target variables across all 3 columns.

INFORMATION AVAILABILITY & TEMPORAL PROTOCOL

Known Future Covariates vs. Historically Observed Covariates

Multi-step forecasting models must respect the arrow of time and real-world information availability. FEV benchmarks and Carla strictly distinguish between features known across the forward prediction window versus realizations available only up to the forecast origin.

FUTURE KNOWN COVARIATESLegitimately available at test time (steps t+1 to t+H)

Scheduled promotions (such as Rossmann's Promo), calendar events (school/state holidays), and operational forecasts (such as Nord Pool's day-ahead grid load and wind power forecasts). These are deterministic or published before market clearing, allowing foundation models to legitimately condition on them across future steps.

HISTORICALLY OBSERVED COVARIATESStrictly restricted to history (time ≤ t)

Stochastic operational observations (such as Rossmann's actual customer footfall Customers, Rohlik's actual precipitation and snowfall, or sudden warehouse shutdowns). Because these are only realized in real-time, future values must never enter a prediction to prevent lookahead target leakage.

Observed Dataset Dynamics: Governed by calendar attributes (day of week, month, payday cycles) and calendar holidays. Series telemetry captures continuous operational volume signals without auxiliary sensor covariates.
ColVariable / Feature NameRoleTypeUnit / ScaleDescription & Business Meaning
CtargetTarget VolumeTARGETNUMcount
Continuous daily operational measurement representing throughput, demand, or activity volume.
AidSeries IdentifierSERIES IDCAT—
Unique categorical entity identifier distinguishing each of the 30 monitored operational series.
BtimestampObservation DateTIMETIME—
Daily calendar timestamp indicating the start of the observation period in ISO 8601 format.
DATASET PREVIEW

Interactive Data Table

Explore rows, feature values, and target labels for Monash High-Density Time Series Benchmark.

Loading dataset...
GLOBAL BENCHMARK PROVENANCE (FEV-BENCH)

Foundation Models vs Traditional ML & Baselines (35 Methods)

Empirical evaluation directly from the AutoGluon fev benchmark suite. Global aggregate ranking: The win rates and skill scores below represent cross-dataset performance across all 100 standardized tasks in fev-bench, rather than results isolated to Monash High-Density Time Series Benchmark.

Benchmark Provenance & Latency Measurement:
100-Task Global Aggregation:The rankings evaluate zero-shot foundation models across 100 heterogeneous benchmark tasks in fev-bench, measuring aggregate generalization across diverse frequencies and prediction horizons.
FEV Python vs. In-Browser WebGPU Timing:The "FEV Latency" column reports server-side Python execution times from the official AutoGluon FEV benchmark. Independently, Carla measures and reports true client-side WebGPU execution inside Google Sheets (247ms on this dataset) with zero server reliance.
#1 Official Ranked Model?ci-proto from Continuous Intelligence ranks #1 on the 35-method official fev-bench view (90.0% SQL win rate; #3 overall at 84.37% across all 52 evaluated models) and can also run locally in our browser WebGPU demo.
ci-proto90.0% Win Rate • 0.53s FEV latency
Evaluated Methods?35 official leaderboard methods (52 total evaluated including closed APIs and checkpoint variants across foundation models, deep neural architectures, trees, and statistical baselines).
35 Official52 total • Continuous Intelligence, H2O.ai, Mistral AI, Google, AWS
Primary Evaluation Metric?Scaled Quantile Loss (SQL) penalizes both point prediction errors and quantile miscalibration across 9 probabilistic levels (p10 to p90).
SQL & MASEProbabilistic calibration
Independent Browser WebGPU?Carla Engine executes Timestral-123M locally in Google Sheets via WebGPU shaders, while ci-proto can also be tested directly in-browser via ONNX Runtime WebGPU and Rust/WASM above.
247ms / windowTimestral-123M (Carla Engine) • Zero server upload
RankMethod / ModelFamilyOrganization
SQL Win Rate?Average win rate (%) against all competing models across the benchmark tasks under Scaled Quantile Loss.
SQL Skill?Normalized skill score relative to the Seasonal Naive baseline (0% = baseline, higher is better).
MASE Win Rate?Average win rate (%) under Mean Absolute Scaled Error.
FEV Latency?FEV benchmark reference latency measured in server Python (median seconds per 100 series). Independent in-browser WebGPU execution for Carla is reported in the telemetry card above.
Link
#1ci-protoFoundation ModelContinuous Intelligence90.0%+51.5%88.4%0.53sRepo ↗
#2TimeH2O-v1Foundation ModelH2O.ai89.7%+50.8%88.2%2.82sRepo ↗
#3Timestral-209MFoundation ModelMistral AI87.1%+50.4%85.4%0.51sRepo ↗
#4t0-gammaFoundation ModelThe Forecasting Company84.7%+49.4%82.0%4.38sRepo ↗
#5Simons-297MFoundation ModelAbel AI Lab82.0%+48.4%80.9%0.94sRepo ↗
#6timesfm-3Foundation ModelGoogle79.3%+49.0%77.6%3.41sRepo ↗
#7iAmTime-base-forecastFoundation ModelWalmart Labs73.7%+46.8%72.5%0.56sRepo ↗
#8Chronos-2Foundation ModelAWS72.2%+47.3%69.9%0.84sRepo ↗
#9t0-betaFoundation ModelThe Forecasting Company69.9%+46.6%66.9%2.29sRepo ↗
#10Toto-2.0-2.5BBaselineCommunity69.4%+44.4%67.8%5.58sRepo ↗
#11TiRex-2Foundation ModelNX-AI69.2%+45.5%66.3%0.27sRepo ↗
#12TiRexFoundation ModelNX-AI62.2%+42.6%58.0%0.24sRepo ↗
#13Toto-2.0-22mBaselineCommunity61.6%+42.9%59.4%0.25sRepo ↗
#14TimesFM-2.5BaselineCommunity58.4%+42.2%56.7%1.87sRepo ↗
#15TabPFN-TS-3Foundation ModelPrior Labs55.5%+43.1%53.3%235sRepo ↗
#16TS-ICLFoundation ModelEDF-Lab55.1%+43.1%53.4%2.49sRepo ↗
#17citras-fmFoundation ModelHitachi52.9%+41.2%50.6%0.60sRepo ↗
#18FlowStateFoundation ModelIBM52.7%+39.9%53.0%2.29sRepo ↗
#19Toto-1.0BaselineCommunity51.6%+40.8%48.8%22.35sRepo ↗
#20TabPFN-TSFoundation ModelPrior Labs50.5%+41.5%48.0%109sRepo ↗
#21Moirai-2.0BaselineCommunity46.2%+39.3%44.2%0.35sRepo ↗
#22Chronos-BoltFoundation ModelAWS45.4%+38.9%42.9%0.26sRepo ↗
#23TFTDeep Learning—34.7%+32.2%31.6%829sRepo ↗
#24Sundial-BaseFoundation ModelTsinghua University30.8%+33.4%37.0%8.01sRepo ↗
#25Stat. EnsembleBaselineCommunity30.6%+20.7%33.6%147sRepo ↗
#26PatchTSTDeep Learning—30.0%+30.1%27.8%664sRepo ↗
#27DeepARDeep Learning—28.9%+29.1%25.9%1113sRepo ↗
#28AutoARIMAStatistical Baseline—26.6%+21.3%24.2%20.14sRepo ↗
#29AutoETSStatistical Baseline—22.5%-26.8%24.1%3.48sRepo ↗
#30CatBoostTree Ensemble—21.8%+23.0%37.5%29.48sRepo ↗
#31LightGBMTree Ensemble—20.3%+21.0%35.7%3.03sRepo ↗
#32AutoThetaStatistical Baseline—17.9%+5.5%22.2%3.28sRepo ↗
#33Seasonal NaiveBaselineCommunity12.0%0.0%13.2%0.48sRepo ↗
#34NaiveStatistical Baseline—7.8%-45.4%12.2%0.47sRepo ↗
#35DriftStatistical Baseline—7.0%-45.8%11.1%0.45sRepo ↗
STATISTICAL DISTRIBUTIONS

Variable Schema & Summary Distributions

Observed numerical ranges, category cardinalities, missing rates, and sample values across 21,900 rows.

ColVariable NameTypeMissingDistinctSummary Stats / DistributionSample Values
CtargetTARGETNUM020,179[0, 4260] μ=563.6 σ=639.71876.3437523922462.90625
AidCAT030—000
BtimestampCAT0730—2018-01-01T00:00:002018-01-02T00:00:002018-01-03T00:00:00
FREQUENTLY ASKED QUESTIONS

Frequently Asked Questions: Monash High-Density Time Series Benchmark

Common questions regarding the Monash High-Density Time Series Benchmark dataset, machine learning task formulations, and in-browser tabular inference.

What is the forecasting horizon for Monash High-Density Time Series Benchmark?

The standardized evaluation horizon is 28 days (4 weeks) evaluated at daily cadence, designed to mirror monthly operational and supply chain replenishment cycles.

Can I run zero-server forecasts on this dataset in Google Sheets?

Yes, Carla enables zero-server in-context time series forecasting directly in spreadsheets using Mistral's Timestral foundation model without requiring Python runtimes or external infrastructure.

Which models perform best on Monash High-Density Time Series Benchmark?

Zero-shot foundation models such as Timestral, TimesFM, and MOIRAI achieve strong generalized accuracy alongside tuned gradient boosted trees (LightGBM), significantly outperforming uncalibrated statistical baselines like Seasonal Naive.

LIVE EVALUATION IN GOOGLE SHEETS

Forecast with Carla: Monash High-Density Time Series Benchmark

Open the pre-loaded Google Sheet and let Carla generate multi-step probabilistic forecasts using Mistral's Timestral directly inside your browser with zero server uploads.

Step 1

Launch Carla

Click once to open the spreadsheet and Carla side panel together.

Step 2

Review the Setup

Carla auto-detects the target sequence, frequency cadence, and temporal covariates.

Step 3

Forecast & Calibrate

Generate multi-step quantile forecasts with zero server uploads.

Provenance & Attribution

Godahewa, R., Bergmeir, C., Webb, G. I., Hyndman, R. J., & Montero-Manso, P. (2021). Monash Time Series Forecasting Repository. NeurIPS 2021 Track on Datasets and Benchmarks.

License: Open AccessData Source: autogluon/fev_datasetsOriginal Source: https://arxiv.org/abs/2105.06643