Benchmark Results

Note

We will continue to update the following benchmark results. If you have proposed new (classic or data-driven) OLPS models, you can send us your paper/code link via Contact Us or raise a pull request. We will add them to this repository and update the leaderboard as soon as possible.

Benchmark Results on Profitability

See also

The benchmark results on profitability and the analysis of the results are available in the paper.

Benchmark Results on Risk Resilience

Besides the benchmark results of profitability, we further evaluate 2 vital risk metrics: Volatility Risk (VR) and Maximum Drawdown (MDD) on FinOL. Table below reveals that:

  1. The risk profiles vary markedly across different OLPS methods. Methods like SSPO and PPT exhibit higher volatility due to their aggressive betting behaviors. In contrast, methods such as UP and ANTI produce lower risk.

  2. For Pattern-matching methods, AICTR and KTPT undertake more risk exposures for pursuing higher returns.

  3. Most Meta-learning methods exhibit moderate risk profiles comparable to bas. This validates their strength in balancing between return and risk by combining multiple sub-methods.

Benchmark Results of VR and MDD on FinOL

Model

NYSE(O)

NYSE(N)

DJIA

SP500

TSE

SSE

HSI

CMEG

CRYPTO

VR

MDD

VR

MDD

VR

MDD

VR

MDD

VR

MDD

VR

MDD

VR

MDD

VR

MDD

VR

MDD

Classic OLPS: Benchmark baseline

Market

0.15

0.21

0.26

0.55

0.28

0.18

0.23

0.28

0.17

0.26

0.18

0.29

0.18

0.36

0.16

0.24

0.41

0.36

Best

0.19

0.17

1.01

0.96

0.45

0.15

0.48

0.16

0.55

0.38

0.49

0.35

0.53

0.60

0.45

0.52

0.54

0.36

UCRP

0.15

0.20

0.27

0.54

0.28

0.18

0.24

0.29

0.17

0.26

0.17

0.23

0.18

0.36

0.13

0.15

0.40

0.34

BCRP

0.18

0.15

0.66

0.77

0.44

0.16

0.45

0.17

0.55

0.38

0.49

0.35

0.34

0.37

0.45

0.51

0.50

0.30

Classic OLPS: Follow-the-winner

UP

0.15

0.20

0.27

0.54

0.28

0.18

0.24

0.29

0.17

0.26

0.17

0.23

0.18

0.36

0.13

0.15

0.40

0.34

EG

0.15

0.20

0.27

0.54

0.28

0.18

0.24

0.29

0.17

0.26

0.17

0.23

0.18

0.36

0.13

0.15

0.40

0.34

SCRP

0.27

0.58

0.72

0.92

0.38

0.23

0.47

0.40

0.56

0.70

0.48

0.64

0.55

0.63

0.52

0.70

1.04

0.89

PPT

0.36

0.39

1.04

0.84

0.44

0.12

1.55

0.60

0.94

0.22

0.41

0.53

0.77

0.77

0.49

0.61

0.92

0.85

SSPO

0.36

0.39

0.92

0.85

0.44

0.12

1.52

0.60

0.95

0.23

0.42

0.57

0.78

0.75

0.52

0.68

0.92

0.85

Classic OLPS: Follow-the-loser

ANTI1

0.21

0.20

0.60

0.62

0.31

0.20

0.74

0.31

0.45

0.17

0.24

0.24

0.49

0.65

0.28

0.37

0.43

0.28

ANTI2

0.24

0.21

0.70

0.62

0.34

0.20

0.90

0.42

0.59

0.16

0.26

0.23

0.62

0.72

0.33

0.46

0.44

0.25

PAMR

0.31

0.70

0.58

0.58

0.40

0.19

0.78

0.48

0.66

0.25

0.32

0.56

0.46

0.69

0.36

0.52

0.70

0.86

CWMR-Var

0.31

0.70

0.62

0.65

0.40

0.19

0.85

0.49

0.66

0.26

0.31

0.56

0.47

0.72

0.37

0.51

0.67

0.79

CWMR-Stdev

0.31

0.70

0.63

0.67

0.40

0.19

0.85

0.49

0.66

0.26

0.31

0.56

0.47

0.72

0.37

0.51

0.67

0.79

OLMAR-S

0.35

0.46

0.94

0.85

0.46

0.16

1.17

0.71

0.71

0.20

0.41

0.58

0.54

0.81

0.46

0.60

0.83

0.75

OLMAR-E

0.35

0.64

1.00

0.86

0.46

0.23

1.24

0.85

0.75

0.28

0.41

0.67

0.55

0.81

0.44

0.64

0.72

0.89

RMR

0.35

0.54

0.90

0.84

0.46

0.21

1.16

0.76

0.72

0.19

0.40

0.55

0.56

0.79

0.44

0.57

0.82

0.74

RPRT

0.35

0.60

1.08

0.87

0.47

0.29

1.65

0.80

0.95

0.30

0.44

0.58

0.56

0.76

0.47

0.59

0.72

0.89

Classic OLPS: Pattern-matching

AICTR

0.36

0.52

1.05

0.93

0.47

0.20

1.35

0.75

0.93

0.31

0.45

0.71

0.56

0.79

0.46

0.62

0.90

0.65

KTPT

0.36

0.64

0.98

0.68

0.43

0.12

1.32

0.64

0.88

0.24

0.38

0.68

0.58

0.78

0.50

0.73

0.79

0.94

Classic OLPS: Meta-learning

SP

0.15

0.20

0.27

0.54

0.28

0.18

0.24

0.29

0.17

0.26

0.17

0.23

0.18

0.36

0.13

0.15

0.40

0.34

ONS

0.19

0.24

0.77

0.89

0.36

0.23

1.22

0.72

0.37

0.28

0.21

0.24

0.60

0.71

0.07

0.14

0.43

0.31

GRW

0.15

0.20

0.26

0.52

0.27

0.15

0.28

0.27

0.31

0.34

0.22

0.27

0.20

0.37

0.15

0.15

0.42

0.46

WAAS

0.15

0.20

0.27

0.54

0.28

0.18

0.24

0.29

0.17

0.26

0.17

0.23

0.18

0.35

0.13

0.16

0.40

0.34

CW-OGD

0.15

0.21

0.37

0.66

0.24

0.16

0.26

0.37

0.21

0.25

0.24

0.33

0.23

0.37

0.24

0.29

0.42

0.34

The analysis reveals that while certain OLPS methods can gain higher profitability, their improvement often accompanies greater risk exposures. Hence, an ideal OLPS method should maximize returns while controlling the downside risk. The benchmark results motivate the development of innovative mechanisms to manage risk in the data-driven OLPS field, such as adaptive position sizing, robust rebalancing rules, and risk-sensitive loss functions.

Benchmark Results on Practicality

We also examine 2 practical OLPS metrics: Average Turnover (ATO) and Running Time (RT). Table below presents the ATO results on FinOL. Several key observations can be summarized:

  1. Methods like UP and EG generate relatively low ATOs by stable rebalancing.

  2. However, some methods that fall into the Follow-the-loser classification (e.g., CWMR and OLMAR) exhibit high ATOs due to their frequent portfolio adjustment, which may raise practical concerns about high transaction costs.

  3. Most Meta-learning methods maintain ATO levels comparable to UP, which are lower than methods like PAMR with higher turnover rates. This demonstrates Meta-learning methods can effectively balance adaptability and trading frequencies.

Benchmark Results of ATO on FinOL

Model

NYSE(O)

NYSE(N)

DJIA

SP500

TSE

SSE

HSI

CMEG

CRYPTO

Classic OLPS: Benchmark baseline

Market

Best

UCRP

0.57%

0.75%

1.24%

1.01%

1.06%

1.98%

1.94%

1.52%

0.96%

BCRP

0.20%

1.15%

0.70%

0.41%

0.30%

0.37%

2.59%

0.44%

1.15%

Classic OLPS: Follow-the-winner

UP

1.37%

0.75%

1.97%

1.01%

2.02%

2.50%

2.66%

2.10%

1.88%

EG

0.55%

0.71%

1.20%

0.96%

1.02%

1.90%

1.86%

1.47%

0.92%

SCRP

7.20%

2.83%

19.36%

12.31%

12.42%

15.45%

14.01%

20.43%

12.63%

PPT

55.43%

54.27%

53.66%

52.18%

55.37%

54.74%

50.88%

51.21%

49.88%

SSPO

52.40%

53.74%

53.46%

52.43%

52.97%

52.55%

48.35%

45.92%

47.11%

Classic OLPS: Follow-the-loser

ANTI1

18.16%

22.99%

13.82%

21.42%

20.47%

16.49%

17.08%

14.90%

16.78%

ANTI2

27.21%

27.40%

20.65%

24.09%

25.51%

25.93%

19.19%

21.25%

25.79%

PAMR

84.99%

67.16%

86.00%

66.67%

78.49%

66.67%

68.28%

74.41%

65.21%

CWMR-Var

84.06%

65.77%

84.85%

65.69%

77.52%

64.42%

67.38%

73.44%

63.70%

CWMR-Stdev

84.06%

65.55%

84.85%

65.59%

77.51%

64.34%

67.32%

73.50%

63.71%

OLMAR-S

64.43%

58.68%

60.99%

56.49%

63.65%

57.63%

60.71%

56.13%

55.64%

OLMAR-E

75.44%

73.34%

71.47%

72.21%

72.83%

69.81%

70.84%

67.00%

62.46%

RMR

66.00%

54.58%

63.00%

54.73%

67.49%

57.97%

58.47%

57.75%

54.69%

RPRT

72.71%

69.78%

65.47%

68.52%

68.70%

66.52%

66.28%

65.33%

58.06%

Classic OLPS: Pattern-matching

AICTR

63.30%

69.44%

58.47%

69.34%

66.41%

63.36%

69.31%

61.37%

59.14%

KTPT

87.05%

81.88%

79.46%

72.07%

76.36%

70.96%

70.94%

63.17%

69.87%

Classic OLPS: Meta-learning

SP

0.57%

0.75%

1.24%

1.01%

1.06%

1.98%

1.94%

1.52%

0.96%

ONS

4.25%

4.66%

9.17%

9.27%

9.53%

7.99%

10.83%

5.12%

6.01%

GRW

0.57%

6.39%

13.24%

14.76%

12.99%

11.55%

11.69%

8.23%

6.07%

WAAS

0.59%

0.76%

1.63%

1.12%

1.22%

2.14%

2.11%

1.76%

1.01%

CW-OGD

0.57%

2.07%

1.96%

6.29%

1.87%

3.59%

4.69%

2.21%

1.21%

Note

“–” indicates that the method incurs (almost) no transaction costs.

The results reveal a trade-off between profitability and practicality. Some methods gain high returns at the cost of heavy portfolio adjustments. Hence, an ideal data-driven OLPS should optimize returns under real-world constraints. The benchmark motivates developing mechanisms to lower rebalancing frequencies and reduce turnover.

In addition to the evaluation of ATO, we also examine the Running Time (RT) metric as a practical measure of OLPS performance. Table below presents the RT results on the FinOL. Several key observations can be summarized:

  1. Some OLPS methods have poor scalability with respect to the number of assets, particularly SCRP and SSPO. This is mainly because their past development relied heavily on the OLPS database, which fortunately does not involve any large-scale datasets.

  2. The majority of OLPS methods perform well in terms of running time.

Benchmark Results of RT on FinOL

Model

NYSE(O)

NYSE(N)

DJIA

SP500

TSE

SSE

HSI

CMEG

CRYPTO

Classic OLPS: Benchmark baseline

Market

Best

UCRP

0.0041322

0.0059764

0.0003916

0.001057

0.000611

0.0002481

0.0002688

0.000459

0.0008224

BCRP

0.0038118

0.0067816

0.0002166

0.001065

0.000996

0.000306

0.0002986

0.000275

0.0008464

Classic OLPS: Follow-the-winner

UP

1.8864056

0.0719632

0.0254685

0.01317

0.091344

0.0412119

0.0512968

0.037949

0.4567986

EG

0.0209176

0.7080021

0.0007138

0.021471

0.0023

0.0009838

0.0014511

0.001174

0.0048522

SCRP

21.3714095

3563.216028

2.1666272

751.7104

5.162115

1.3369282

4.1774117

1.326479

10.0444827

PPT

0.0220539

1.4248651

0.000765

0.067611

0.002196

0.0011361

0.0016884

0.000959

0.0038207

SSPO

10.365433

1593.101725

4.6482987

285.346

11.55126

5.7671978

6.5936851

2.656706

19.6497393

Classic OLPS: Follow-the-loser

ANTI1

2.412872

942.9277071

0.9028023

89.10436

4.825674

1.5139913

3.4485262

1.232613

6.3589876

ANTI2

5.3055469

1019.015005

0.7316441

67.97805

6.081596

1.8371354

5.6760943

2.697906

11.1037912

PAMR

0.0281713

0.6058916

0.0016882

0.016369

0.003871

0.0021293

0.0024729

0.002359

0.0094379

CWMR-Var

0.0486561

5.0129114

0.0021516

0.384601

0.006705

0.0036254

0.0053173

0.002194

0.0118123

CWMR-Stdev

0.0554241

7.6208348

0.0030268

0.89136

0.009707

0.0057918

0.0093459

0.00356

0.0203114

OLMAR-S

0.0388321

0.6172465

0.0016541

0.012853

0.003063

0.0014739

0.0017771

0.001449

0.0065995

OLMAR-E

0.0398049

0.6097762

0.0012705

0.012054

0.002877

0.0018873

0.0021814

0.001494

0.0063896

RMR

0.1065905

0.4709969

0.005274

0.060043

0.012206

0.0084572

0.0093933

0.009332

0.0335466

RPRT

0.0483275

2.3619424

0.0022602

0.089113

0.003629

0.0017498

0.0022784

0.001699

0.0074966

Classic OLPS: Pattern-matching

AICTR

0.1552033

10.3116284

0.0036974

0.441138

0.009328

0.0039644

0.0060411

0.003575

0.0364883

KTPT

1.33071

5.1376433

0.0563649

0.506718

0.498936

0.6813439

0.5995139

0.704583

1.9678092

Classic OLPS: Meta-learning

SP

0.0786008

1.8743168

0.0124197

0.336125

0.016422

0.0049131

0.0063926

0.016176

0.0561672

ONS

1.5128513

701.8432832

0.1277076

13.5781

0.250247

0.0999844

0.1304165

0.094656

0.5413568

GRW

0.7999147

44.0762048

0.0620426

8.250292

0.374519

0.0961017

0.3086175

0.085338

0.8238871

WAAS

1.0277107

295.7563543

0.1262029

45.77576

0.626888

0.425847

0.8555393

0.197126

1.9221044

CW-OGD

0.0373499

1.8004252

0.0027351

0.210646

0.007928

0.0042008

0.0057244

0.003265

0.0165171

Note

All time are reported in seconds.

RT is an important consideration that is often overlooked in backtesting but critical for real-world OLPS task. To address this challenge, FinOL offers researchers access to two large datasets: NYSE(N) and SP500. These datasets provide researchers with opportunities to develop OLPS methods that are not only profitable but also practical.