Dataset Viewer
Auto-converted to Parquet Duplicate
system
stringclasses
6 values
run
stringclasses
4 values
day
stringdate
2025-04-02 00:00:00
2025-04-04 00:00:00
rmse
float64
0.75
47.1
โŒ€
mape
float64
0.41
29.6
โŒ€
dtw
float64
0.1
29.5
โŒ€
dtw_eq7
float64
6.3
930
โŒ€
dtw_eq7n
float64
0.01
1.53
โŒ€
qq
float64
0.71
1
โŒ€
volsim
float64
0
0.93
โŒ€
mars
seed000
2025-04-02
10.466372
6.23424
6.234751
204.276619
0.338256
0.904457
0.537463
mars
seed000
2025-04-03
24.268096
15.014213
15.022127
479.256503
0.769614
0.863794
0.393182
mars
seed000
2025-04-04
28.628343
18.35008
18.345008
565.364474
0.929847
0.816642
0.401803
mars
seed001
2025-04-02
18.725571
9.600504
9.599485
368.312016
0.609878
0.7059
0.192359
mars
seed001
2025-04-03
24.228447
14.991744
14.986509
478.473498
0.768357
0.865281
0.287208
mars
seed001
2025-04-04
47.10606
29.576724
29.521772
929.929463
1.529443
0.808102
0.0731
mars
seed002
2025-04-02
5.916648
3.474789
3.474395
110.528358
0.183021
0.899188
0.314514
mars
seed002
2025-04-03
0.747789
0.406025
0.149238
10.151654
0.016302
0.720146
0.40436
mars
seed002
2025-04-04
1.071567
0.536699
0.10482
6.300675
0.010363
0.884331
0.579532
marketsim
seed000
2025-04-02
3.330931
2.067923
2.068344
65.780626
0.108924
0.987454
0.75501
marketsim
seed000
2025-04-03
11.486787
7.118674
7.122159
226.845866
0.364281
0.996265
0.569997
marketsim
seed000
2025-04-04
11.444663
7.291418
7.299193
226.013979
0.371722
0.994902
0.754649
marketsim
seed001
2025-04-02
4.581111
2.946591
2.947
90.46969
0.149806
0.956001
0.800512
marketsim
seed001
2025-04-03
10.567912
6.581516
6.583541
208.699546
0.33514
0.988222
0.429279
marketsim
seed001
2025-04-04
8.8322
5.585149
5.59367
167.331685
0.275208
0.990064
0.726627
marketsim
seed002
2025-04-02
4.9269
3.171518
3.1719
97.298483
0.161114
0.980552
0.865881
marketsim
seed002
2025-04-03
10.79179
6.724491
6.726911
213.120779
0.34224
0.979808
0.928768
marketsim
seed002
2025-04-04
11.031684
7.048399
7.053313
217.858305
0.358309
0.988795
0.518492
twinmarket
seed000
2025-04-02
4.534743
2.890676
1.921886
83.314202
0.137958
0.931376
0.5962
twinmarket
seed000
2025-04-03
9.598904
5.993442
5.995302
189.563159
0.30441
0.983274
0.52636
twinmarket
seed000
2025-04-04
6.092772
3.708422
3.721803
120.322598
0.197893
0.985151
0.576279
twinmarket
seed001
2025-04-02
4.766111
3.061715
3.062116
94.123148
0.155856
0.979323
0.456132
twinmarket
seed001
2025-04-03
9.621284
6.007022
6.008855
190.005129
0.30512
0.94992
0.694637
twinmarket
seed001
2025-04-04
6.051492
3.687728
3.700866
113.160771
0.186114
0.985684
0.609116
twinmarket
seed002
2025-04-02
4.741922
3.055414
3.055807
93.645455
0.155065
0.97262
0.513993
twinmarket
seed002
2025-04-03
9.587606
5.98526
5.987176
189.340054
0.304052
0.985889
0.522477
twinmarket
seed002
2025-04-04
6.074305
3.700426
3.713661
119.957918
0.197293
0.984216
0.557269
background
seed000
2025-04-02
4.735679
3.050844
3.051253
93.522174
0.154861
0.968826
0.4876
background
seed000
2025-04-03
9.596818
5.991714
5.993599
189.521965
0.304344
0.984552
0.484268
background
seed000
2025-04-04
6.074976
3.702164
3.715355
119.971171
0.197315
0.985156
0.494071
background
seed001
2025-04-02
4.732632
3.049079
3.049482
93.461984
0.154761
0.974314
0.502906
background
seed001
2025-04-03
9.583219
5.983426
5.985297
189.253417
0.303913
0.978211
0.499418
background
seed001
2025-04-04
6.053595
3.687669
3.700866
119.548925
0.196621
0.98566
0.515333
background
seed002
2025-04-02
4.736118
3.051434
3.051833
93.530834
0.154875
0.96523
0.530128
background
seed002
2025-04-03
9.583988
5.984197
5.986052
189.268604
0.303937
0.984968
0.475098
background
seed002
2025-04-04
6.067098
3.698147
3.711293
119.815579
0.197059
0.983082
0.502862
flat prev close
-
2025-04-02
2.866584
1.839456
1.839867
56.610507
0.09374
null
0
flat prev close
-
2025-04-03
4.390559
2.711433
2.713376
86.706597
0.139238
null
0
flat prev close
-
2025-04-04
4.369642
2.559104
2.545062
86.293509
0.141926
null
0
real JNJ
-
2025-04-02
null
null
null
null
null
null
null
real JNJ
-
2025-04-03
null
null
null
null
null
null
null
real JNJ
-
2025-04-04
null
null
null
null
null
null
null

FinanceMarketSimulation: evaluation results

The result files (CSV, JSON) of every evaluation report of FinanceMarketSimulation (GitHub: DA2I2-SLM/FinanceMarketSimulation, branch benchmark). The reports themselves (figures, every number as tables, how to reproduce) are in the repository's docs/evaluation/, one folder per report. This dataset has the same layout, so one command restores every file to its place in a clone:

hf download hiepnh137/FinanceMarketSimulation-bench --repo-type dataset --local-dir docs/evaluation
Folder Report Files
bench/ the benchmark suite: MarS's order model, MarketSim and TwinMarket (plus a background-only control) on the MarS and MarketSim benchmarks, one LLM (gpt-oss-120b) 20
mars_experiments/ experiments 1, 2, 3a, 3b: MarS's realism experiments (ICLR 2025) on the repository's own populations 7
hyperliquid/ experiment 4: a real crypto market (Hyperliquid) with participant identity 8
marketsim/ experiment 5: MarketSim on JNJ against real JNJ (tariff days), with six common baselines 11
mars_checkpoint/ MarS's released order model as a baseline on the MarketSim and MarS benchmarks 10
twinmarket/ TwinMarket as a baseline on the MarketSim and MarS benchmarks 11
mars_reference/ MarS's own published numbers and the outputs of MarS's report_stylized_facts.py 11

The benchmark suite at a glance

Benchmark Measure MarS (order model) MarketSim (LLM) TwinMarket + bg Background only Reference
MarS Cont facts as the real windows (of 11) 10 7 6 5 11 (MarS published)
MarS Overlap, 1-min returns (โ†‘) 0.78 0.74 0.62 0.44 0.95
MarS Overlap, window volatility (โ†‘) 0.70 0.54 0.40 0.30 0.86
MarS W1, 1-min returns (โ†“, ร—1e-4) 3.9 19.6 6.7 7.0 1.4
MarS Days closing at the ยฑ10 % limit 13 / 64 46 / 62 0 / 64 0 / 64
JNJ MAPE %, 3 days (โ†“) 10.91 5.39 4.23 4.24 2.37 (flat previous close)
JNJ VolSim, 3 days (โ†‘) 0.35 0.71 0.56 0.50
JNJ Overlap of 1-min returns with real JNJ (โ†‘) 0.61 0.77 0.41 0.40
JNJ Cont facts as real JNJ (of 11) 11 10 10 9

Between the methods (no real data has these; the same numbers on both benchmarks are in docs/evaluation/bench/tables.md in the repository):

Benchmark Measure MarS (order model) MarketSim (LLM) TwinMarket + bg Background only
MarS In-context accuracy, mean of 12 measures (chance 0.33) 0.56 0.67 0.47 0.37
JNJ In-context accuracy, mean of 12 measures 0.54 0.67 0.60 0.55
MarS Impact exponent delta, fit on the mean peak (square-root law: 0.5) 0.43 [-0.21, 1.44] 0.65 [0.35, 1.00] 0.49 [0.21, 1.02] 1.56 [0.65, 1.67]
JNJ Impact exponent delta, fit on the mean peak 0.79 [0.37, 1.19] 0.08 [-0.21, 0.32] 0.94 [0.87, 1.04] 0.60 [0.25, 0.80]
MarS Order flow: overlap of a method's two halves of seeds (its noise floor) 0.89 0.84 0.94 0.95
JNJ Order flow: overlap of a method's two halves of seeds 0.82 0.72 0.86 0.85
both Participant-level measures n/a (pooled orders) yes yes yes

Files

The benchmark suite (bench/)

There is one file per benchmark and table, not one per framework. The prefix names the benchmark: mars_* is the MarS benchmark and marketsim_* the MarketSim (JNJ) benchmark. Each file holds the rows of every method: MarS's order model, MarketSim, TwinMarket and the background, plus the real data or references where the table has them. One column names the method:

File Rows Method column Content
runs.csv 28 method one row per run of both benchmarks: days, scored days, hours, LLM calls and errors
mars_facts.csv 66 dataset MarS benchmark: Cont's 11 facts for the real windows, MarS's published simulation and each method
mars_similarity.csv 90 dataset MarS benchmark: overlap, KS and Wasserstein-1 against the real windows of all 500 stocks
mars_similarity_band.csv 90 dataset the same against the 350 real stocks priced 5-20 CNY
mars_sizes.csv 7 dataset windows and groups scored per dataset
mars_days.csv 254 method every simulated day: close, return, volume, trades, ยฑ10 % limit touched or closed at
marketsim_cells.csv 42 system JNJ benchmark: the MarketSim paper's path metrics per run and day (RMSE, MAPE, DTW, Q-Q, VolSim)
marketsim_qualitative.csv 14 system JNJ: the paper's qualitative rows per run (kurtosis, GARCH, ADF, ...)
marketsim_facts.csv 55 source JNJ: Cont's 11 facts per method and real JNJ
marketsim_fact_cells.csv 429 source JNJ: Cont's facts per run and day
marketsim_days.csv 36 method every simulated JNJ day: close, return, volume, trades
marketsim_reference.csv 912 dataset JNJ: every metric of the metric reference, one dataset per run, with the reason where a metric cannot run
mars_flow.csv, marketsim_flow.csv 120 each dataset, ref order flow (spread, order volume, aggressive and buy ratio): overlap, KS and W1 for every pair of methods (mars vs marketsim) and for each method's two halves of seeds (mars[a] vs mars[b])
mars_lob.csv, marketsim_lob.csv 420 each dataset, ref LOB-Bench (L1 and W1 per score) on the first scored day: seed 0 of every pair of methods, and seed 1 against seed 0 of a method (mars/seed000 vs mars/seed001)
mars_participants.csv, marketsim_participants.csv 272, 204 dataset participant-level measures of every run on the first scored day (twinmarket/seed000)
mars_counterfactual.csv, marketsim_counterfactual.csv 192 each dataset in-context accuracy (accuracy with CI, single rollout, shuffled baseline per measure) and market impact (peak and permanent bp per TWAP size, the square-root fit) of each method's paired tasks (mars)

The labels each framework has in the method column:

Framework method / system / source dataset (MarS benchmark) dataset (marketsim_reference.csv) runs.csv
MarS (order model) mars mars_runs mars/seed000 .. 002 MarS (order model 10M)
MarketSim (LLM managers) marketsim marketsim_runs marketsim/seed000 .. 002 MarketSim (gpt-oss-120b managers)
TwinMarket + background twinmarket twinmarket_runs twinmarket/seed000 .. 002 TwinMarket (gpt-oss-120b) + background
Background only background background_runs background/seed000 .. 002 Background only
References real JNJ, flat prev close mars (real windows), mars_5_20, mars_published jnj

For example, TwinMarket's JNJ path metrics are the rows of marketsim_cells.csv with system == "twinmarket". Its MarS-benchmark overlap is in the rows of mars_similarity.csv with dataset == "twinmarket_runs".

mars_experiments/

File Rows
mars_experiments/e1_fact_summary.csv 66
mars_experiments/e1_sample_sizes.csv 6
mars_experiments/e2_impact_fit.json JSON
mars_experiments/e2_impact_summary.csv 10
mars_experiments/e3_table7.csv 10
mars_experiments/e3a_distribution_similarity.csv 90
mars_experiments/e3b_accuracy.csv 48

hyperliquid/

File Rows
hyperliquid/e4_fact_summary.csv 55
hyperliquid/e4_market_similarity.csv 24
hyperliquid/e4_noise_floor.csv 22
hyperliquid/e4_noise_ladder.csv 6
hyperliquid/e4_participants.csv 88
hyperliquid/e4_participants_by_coin.csv 352
hyperliquid/e4_participants_vs_reference.csv 68
hyperliquid/e4_sample_sizes.csv 5

marketsim/

File Rows
marketsim/marketsim_cells.csv 124
marketsim/marketsim_combined.csv 13
marketsim/marketsim_cont_facts.csv 187
marketsim/marketsim_cont_facts_cells.csv 1331
marketsim/marketsim_garch_fit.csv 1
marketsim/marketsim_lob_bench.csv 63
marketsim/marketsim_main_3days.csv 15
marketsim/marketsim_main_day1.csv 17
marketsim/marketsim_market_rows.csv 11
marketsim/marketsim_paper_layout.csv 30
marketsim/marketsim_qualitative_cells.csv 44

mars_checkpoint/

File Rows
mars_checkpoint/mars_checkpoint_bench_jnj_cells.csv 3
mars_checkpoint/mars_checkpoint_bench_jnj_cont_facts.csv 22
mars_checkpoint/mars_checkpoint_bench_jnj_cont_facts_cells.csv 66
mars_checkpoint/mars_checkpoint_bench_jnj_market_rows.csv 3
mars_checkpoint/mars_checkpoint_bench_jnj_qualitative_cells.csv 3
mars_checkpoint/mars_checkpoint_bench_mars_facts.csv 77
mars_checkpoint/mars_checkpoint_bench_mars_market_rows.csv 16
mars_checkpoint/mars_checkpoint_bench_mars_sample_sizes.csv 7
mars_checkpoint/mars_checkpoint_bench_mars_similarity.csv 72
mars_checkpoint/mars_checkpoint_bench_reference.csv 164

twinmarket/

File Rows
twinmarket/twinmarket_bench_activity.csv 7
twinmarket/twinmarket_bench_jnj_cells.csv 18
twinmarket/twinmarket_bench_jnj_cont_facts.csv 33
twinmarket/twinmarket_bench_jnj_cont_facts_cells.csv 231
twinmarket/twinmarket_bench_jnj_market_rows.csv 18
twinmarket/twinmarket_bench_jnj_qualitative_cells.csv 8
twinmarket/twinmarket_bench_mars_facts.csv 55
twinmarket/twinmarket_bench_mars_sample_sizes.csv 5
twinmarket/twinmarket_bench_mars_similarity.csv 48
twinmarket/twinmarket_bench_reference.csv 885
twinmarket/twinmarket_bench_vs_marketsim_cells.csv 142

mars_reference/

File Rows
mars_reference/paper_table7.csv 10
mars_reference/stylized_facts/cont1.csv 10000
mars_reference/stylized_facts/cont10.csv 20000
mars_reference/stylized_facts/cont11.csv 21000
mars_reference/stylized_facts/cont2.csv 20000
mars_reference/stylized_facts/cont3.csv 20000
mars_reference/stylized_facts/cont5.csv 1000
mars_reference/stylized_facts/cont6.csv 20000
mars_reference/stylized_facts/cont7.csv 20000
mars_reference/stylized_facts/cont9.csv 20000
mars_reference/stylized_facts_presence.csv 11

Where the header is dataset, ref, metric, level, key, name, value, ci_low, ci_high, note, status, a file is the repository's tidy metric table: one row per dataset, metric, key and statistic, with status giving the reason when a metric cannot run. What each file holds is described in the README of its folder in the repository.

Downloads last month
230