Financial Time Series Forecasting Repository (FTSFR)

Financial Time Series Forecasting Repository (FTSFR)#

A standardized macro-finance data repository for global time series forecasting research

GitHub

Objectives#

The Financial Time Series Forecasting Repository (FTSFR) addresses a critical gap in quantitative finance research: the lack of standardized datasets for evaluating and comparing time series forecasting models. When researchers evaluate their own forecasting models on their own arbitrarily chosen datasets, apples-to-apples comparisons across forecasting algorithms are not possible.

This repository provides a comprehensive collection of financial and macroeconomic datasets that have been cleaned, formatted, and standardized according to academic best practices. We hope that this repository will serve as a useful time series forecasting benchmark for the research community.

Key objectives:

  • Provide apples-to-apples comparisons for time series forecasting algorithms

  • Ensure datasets are prepared using consistent methodology

  • Follow academic standards from published research

  • Enable reproducible forecasting benchmarks across financial domains

We are inspired by the work of the Monash Time Series Forecasting Repository, which provides freely available datasets for benchmarking time series forecasting models. However, we focus on datasets from finance and economics, which often require subscriptions to access but are commonly available at universities through services like WRDS (Wharton Research Data Services).

Our Solution#

Although most of the data in this repository requires a subscription to access, we provide a codebase that automates the end-to-end process for assembling various financial and macroeconomic data, from the data pull step to the final cleaning and transformation step. Our quickstart guide provides a simple guide to get you started.

Datasets#

Our benchmark includes 25 datasets spanning multiple financial domains. Each dataset follows cleaning procedures from published academic papers, ensuring that benchmarks reflect established best practices.

Dataset Overview#

Dataset Name

Description

Citation

Returns Data

CDS Contract

Monthly returns for individual CDS contracts (follows Palhares 2012)

Palhares (2012)

CDS Portfolio

20 CDS portfolios by tenor and credit quality

He, Kelly, Manela (2017)

Commodity

Monthly returns for commodity futures

Yang (2013)

Corporate Bond

Monthly returns for individual corporate bonds from TRACE

Dickerson et al. (2024); Nozawa (2017)

Corporate Portfolio

Monthly returns for corporate bond portfolios by credit spread

Nozawa (2017)

CRSP Stock

Monthly stock returns from CRSP database

Fama & French (1993)

FF25 Size-BM

Daily Fama-French 25 portfolios: size and book-to-market

Fama & French (1993)

FX

Daily foreign exchange returns vs USD

Lettau et al. (2014)

SPX Options Portfolios

Monthly returns for individual SPX option contracts

Constantinides et al. (2013)

Treasury Bond

Monthly returns for individual Treasury bonds from CRSP

Gürkaynak et al. (2007)

Treasury Portfolio

Monthly returns for Treasury bond portfolios by maturity

Gürkaynak et al. (2007)

Basis Spread Data

CDS-Bond

Monthly CDS-bond basis spreads

Siriwardane et al. (2022)

CIP

Monthly covered interest parity deviations

Du et al. (2018)

TIPS-Treasury

Monthly TIPS-Treasury basis spreads

Fleckenstein et al. (2014)

Treasury-SF

Monthly Treasury-SF arbitrage spreads

Jermann (2020)

Treasury-Swap

Monthly Treasury-Swap arbitrage spreads

Siriwardane et al. (2022)

Other Financial Data

Bank Cash Liquidity

Quarterly cash liquidity from call report data

Drechsler et al. (2017)

Bank Leverage

Quarterly leverage ratios from call report data

Drechsler et al. (2017)

BHC Cash Liquidity

Quarterly bank holding company cash liquidity

Drechsler et al. (2017)

BHC Leverage

Quarterly bank holding company leverage ratios

Drechsler et al. (2017)

HKM Daily Factor

Intermediary risk factors (capital ratio, risk factor, returns, leverage)

He, Kelly, Manela (2017)

HKM Monthly Factor

Same as above, but monthly

He, Kelly, Manela (2017)

Treasury Yield Curve

Daily Nelson-Siegel-Svensson zero-coupon yields (1-30 years)

Gürkaynak et al. (2007)

Source: Authors’ analysis

Dataset Statistics#

The following table provides key statistics for each dataset after filtering and cleaning:

Dataset

Frequency

Entities (Before)

Entities (After)

Retention

Median Length (After)

Date Range

Basis Spreads

CDS-Bond

Monthly

3,402

1,516

44.6%

57

2002-09 to 2022-09

CIP

Monthly

8

8

100.0%

272

2001-12 to 2025-02

TIPS-Treasury

Monthly

4

4

100.0%

251

2004-07 to 2025-05

Treasury-SF

Monthly

5

5

100.0%

247

2004-06 to 2025-01

Treasury-Swap

Monthly

7

7

100.0%

207

2001-12 to 2025-08

Returns (Portfolios)

CDS Portfolio

Monthly

20

4

20.0%

276

2001-01 to 2023-12

Corporate Portfolio

Monthly

10

10

100.0%

242

2002-08 to 2022-09

FF25 Size-BM

Daily

25

25

100.0%

36,160

1926-07 to 2025-06

SPX Options Portfolios

Monthly

18

18

100.0%

288

1996-01 to 2019-12

Treasury Portfolio

Monthly

10

10

100.0%

668

1970-01 to 2025-08

Returns (Disaggregated)

CDS Contract

Monthly

6,552

234

3.6%

54

2001-02 to 2023-12

CRSP Stock

Monthly

26,757

25,095

93.8%

94

1926-01 to 2024-12

CRSP Stock (ex-div)

Monthly

26,757

25,095

93.8%

94

1926-01 to 2024-12

Commodity

Monthly

23

23

100.0%

511

1970-01 to 2025-08

Corporate Bond

Monthly

23,473

16,719

71.2%

52

2002-08 to 2022-09

FX

Monthly

9

9

100.0%

276

1999-02 to 2025-02

Treasuries

Monthly

2,054

1,912

93.1%

49

1970-01 to 2025-08

Other

BHC Cash Liquidity

Quarterly

13,770

6,351

46.1%

69

1976-03 to 2020-03

BHC Leverage

Quarterly

13,761

6,653

48.3%

67

1976-03 to 2020-03

Bank Cash Liquidity

Quarterly

23,862

17,383

72.8%

82

1976-03 to 2020-03

Bank Leverage

Quarterly

22,965

17,295

75.3%

82

1976-03 to 2020-03

HKM All Factor

Monthly

4

4

100.0%

516

1970-01 to 2012-12

HKM Daily Factor

Daily

4

4

100.0%

6,918

2000-01 to 2018-12

HKM Monthly Factor

Monthly

4

4

100.0%

587

1970-01 to 2018-11

Treasury Yield Curve

Daily

30

30

100.0%

17,902

1961-06 to 2025-09

Source: Generated automatically from dataset processing pipeline

Forecasting Results#

We provide comprehensive baseline results across all datasets using a diverse set of forecasting models, from classical statistical methods to modern deep learning architectures. Results are evaluated using multiple metrics to provide complementary perspectives on forecasting performance.

Overall Model Performance#

The table below shows median and mean performance statistics across all 25 datasets. MASE (Mean Absolute Scaled Error) is the primary metric used in forecasting literature, with values < 1.0 indicating better performance than a seasonal naive benchmark. Relative MASE compares each model to the Historic Average baseline. R² (out-of-sample R²) measures the percentage reduction in mean squared error relative to predicting the historical average.

Model

N

Med MASE

Mean MASE

Med Rel MASE

Mean Rel MASE

Med R²

Mean R²

NBEATS

25

0.786

0.950

0.477

0.606

0.098

-0.062

NHITS

25

0.806

0.912

0.488

0.611

0.333

-0.077

Theta

25

0.814

0.927

0.505

0.599

0.379

-0.019

DLinear

25

0.829

0.977

0.554

0.606

0.307

0.019

NLinear

25

0.837

0.944

0.518

0.612

0.353

-1.755

ARIMA

25

0.841

0.933

0.498

0.622

0.359

0.278

Transformer

25

0.852

0.998

0.491

0.602

0.245

0.185

KAN

25

0.861

1.063

0.481

0.641

-0.030

0.003

TiDE

25

0.875

0.969

0.586

0.634

0.234

-0.200

SES

25

1.020

1.359

0.638

0.700

0.169

-0.043

DeepAR

25

1.148

1.791

0.851

0.837

-0.081

-4.514

HistAvg

25

1.819

2.957

—

—

0.000

0.000

Note: MASE shows absolute performance (lower is better), Relative MASE shows performance relative to Historic Average (lower is better), and R² shows out-of-sample predictive power (higher is better). Models are sorted by median MASE. Bold indicates best performance in each column.

Sources: Bloomberg, Board of Governors of the Federal Reserve System, Center for Research in Security Prices, U.S. Call Reports, WRDS TRACE, OptionMetrics, S&P Global, Authors’ analysis

Performance by Dataset Category#

Model rankings vary significantly across different types of financial data. The table below disaggregates performance by dataset category:

Basis Spreads

Model

N

Med MASE

Mean MASE

Med Rel MASE

Mean Rel MASE

Med R²

Mean R²

NLinear

5

0.395

0.637

0.299

0.407

0.514

0.651

Theta

5

0.411

0.672

0.297

0.449

0.544

0.529

Transformer

5

0.421

0.723

0.305

0.475

0.581

0.563

ARIMA

5

0.446

0.700

0.306

0.457

0.539

0.588

DLinear

5

0.455

0.836

0.495

0.466

0.431

0.562

NHITS

5

0.464

0.664

0.272

0.446

0.580

0.550

KAN

5

0.465

0.766

0.337

0.497

0.405

0.509

TiDE

5

0.491

0.862

0.311

0.627

0.517

0.185

DeepAR

5

0.567

1.147

0.693

0.615

0.186

0.102

NBEATS

5

0.587

0.685

0.329

0.441

0.706

0.552

SES

5

0.939

1.208

0.601

0.672

0.345

0.250

HistAvg

5

1.784

2.126

—

—

0.000

0.000

Returns

Model

N

Med MASE

Mean MASE

Med Rel MASE

Mean Rel MASE

Med R²

Mean R²

KAN

12

0.803

1.183

0.959

0.893

-0.073

-0.267

Transformer

12

0.815

1.119

0.948

0.840

-0.043

-0.099

NBEATS

12

0.825

1.024

0.955

0.858

-0.059

-0.513

DLinear

12

0.827

0.988

0.974

0.837

-0.029

-0.469

NLinear

12

0.845

1.015

0.992

0.876

-0.184

-4.149

SES

12

0.853

1.501

0.965

0.899

-0.021

-0.437

TiDE

12

0.859

0.975

0.913

0.833

-0.077

-0.742

Theta

12

0.860

1.045

0.965

0.862

-0.015

-0.614

NHITS

12

0.869

1.000

0.996

0.881

-0.156

-0.675

ARIMA

12

0.873

1.035

1.003

0.902

-0.073

-0.009

HistAvg

12

0.873

2.519

—

—

0.000

0.000

DeepAR

12

1.181

2.071

1.007

1.075

-0.125

-7.815

Other (Bank Metrics, Factors, Yield Curve)

Model

N

Med MASE

Mean MASE

Med Rel MASE

Mean Rel MASE

Med R²

Mean R²

NHITS

8

0.796

0.937

0.327

0.308

0.465

0.429

Theta

8

0.811

0.908

0.313

0.298

0.474

0.531

ARIMA

8

0.830

0.926

0.328

0.305

0.481

0.514

DLinear

8

0.839

1.049

0.378

0.346

0.380

0.411

NBEATS

8

0.867

1.004

0.380

0.331

0.459

0.232

Transformer

8

0.888

0.988

0.363

0.326

0.437

0.375

KAN

8

0.912

1.069

0.404

0.353

0.413

0.090

TiDE

8

0.918

1.028

0.362

0.341

0.414

0.372

NLinear

8

0.925

1.031

0.412

0.344

0.398

0.333

SES

8

1.118

1.242

0.491

0.421

0.292

0.364

DeepAR

8

1.359

1.773

0.473

0.619

0.043

-2.446

HistAvg

8

2.653

4.132

—

—

0.000

0.000

Note: Metrics are computed within each dataset category. Lower MASE/Relative MASE values indicate better performance; higher R² values indicate better performance. Bold indicates best performance in each column.

Key Findings#

  • Basis spreads exhibit pronounced seasonal structure and mean reversion, where models like NLinear, NHITS, and NBEATS excel by decomposing series into trend and seasonal components

  • Asset returns are notoriously difficult to forecast, with nearly all models achieving R² values near zero—consistent with decades of empirical asset pricing research showing that the historical mean is hard to beat

  • Other datasets (bank metrics, intermediary factors, yield curves) favor classical approaches like Theta and ARIMA, which balance flexibility with parsimony

Getting Started#

  1. Installation: Follow our quickstart guide to set up the environment

  2. Configuration: Set up your data source credentials and specify available subscriptions

  3. Data Download: Run the automated ETL pipeline to build your standardized datasets

  4. Benchmarking: Use our baseline models to establish performance benchmarks

Documentation#

📚 Documentation

📊 Data Cleaning Procedure Summaries

🔧 Development

Key Features#

  • 📡 Streamlined data collection from multiple sources (WRDS, FRED, Bloomberg, etc.)

  • 🎯 Academic standards: Replicates cleaning procedures from academic literature

  • 🤖 Fully reproducible data preparation workflows

  • 📈 Spans multiple financial domains:

    • Equity Markets: Returns, portfolios, and characteristics

    • Fixed Income: Treasury, corporate, and sovereign bonds

    • Credit Markets: CDS spreads and bond-CDS basis

    • Derivatives: Options and futures data

    • Foreign Exchange: Currency portfolios and rates

    • Commodities: Futures and spot prices

    • Banking: Call report data and bank-specific metrics

Contributing#

We welcome contributions from the research community. Whether you’re:

  • Adding new data sources

  • Improving cleaning methodologies

  • Implementing additional forecasting benchmarks

  • Enhancing documentation

Please see our contribution guidelines for more information.

Acknowledgments#

We would like to thank the following individuals. With their permission, we have adapted and used pieces of their code in this repository:

  • Om Mehta and Kunj Shah, for their replication of the Covered Interest Rate Parity (CIP) arbitrage spreads (Siriwardane et al. 2022; Rime et al. 2017), available at Kunj121/CIP

  • Kyle Parran and Duncan Park, for their replication of the construction of commodity futures returns (He et al. 2017; Yang 2013), available at kyleparran/final_project_group_09

  • Haoshu Wang and Guanyu Chen, for their replication of the Treasury Spot-Futures basis (Siriwardane et al. 2022)

  • Arsh Kumar and Raiden Egbert, for their replication of the Treasury Swap basis (Siriwardane et al. 2022)

  • Bailey Meche and Raul Renteria, for their replication of the TIPS-Treasury basis (Siriwardane et al. 2022)

Project Team#

Project Lead:

  • Jeremiah Bejarano, Office of Financial Research, U.S. Department of the Treasury and Financial Mathematics Program, University of Chicago

Project Collaborators:

  • Viren Desai

  • Kausthub Keshava

  • Arsh Kumar

  • Zixiao Wang

  • Vincent Hanyang Xu

  • Yangge Xu

Views and opinions expressed are those of the authors and do not necessarily represent official positions or policy of the Office of Financial Research (OFR) or the U.S. Department of the Treasury.

Citation#

If you use FTSFR in your research, please cite:

Bejarano, J. et al. (2024). Financial Time Series Forecasting Repository:
A Standardized Macro-Finance Data Repository for Global Time Series Forecasting.
[Working Paper]

Note

Academic Collaboration: This project represents a collaborative effort among researchers to advance the state of financial time series forecasting. We encourage academic partnerships and welcome contributors who are interested in pushing the boundaries of empirical finance research.

Indices and Tables#