Datasets:
date date32 | symbol stringclasses 97
values | open float64 9.81 1.44k | high float64 10.4 1.45k | low float64 8.95 1.41k | close float64 9.17 1.44k | volume int64 0 1.14B | adj float64 0.85 1 | dividend float64 0 5 | split float64 0 25 |
|---|---|---|---|---|---|---|---|---|---|
2024-01-02 | AAPL | 187.15 | 188.44 | 183.89 | 185.64 | 82,488,700 | 0.987955 | 0 | 0 |
2024-01-02 | ABBV | 154.9 | 160.14 | 154.77 | 159.82 | 8,059,900 | 0.905054 | 0 | 0 |
2024-01-02 | ADBE | 589.51 | 590.44 | 576.76 | 580.07 | 2,840,200 | 1 | 0 | 0 |
2024-01-02 | AMD | 144.28 | 144.4 | 137.43 | 138.58 | 64,902,000 | 1 | 0 | 0 |
2024-01-02 | AMZN | 151.54 | 152.38 | 148.39 | 149.93 | 47,339,400 | 1 | 0 | 0 |
2024-01-02 | AVGO | 109.212 | 110.2 | 107.709 | 108.538 | 28,831,000 | 0.972514 | 0 | 0 |
2024-01-02 | BAC | 33.39 | 34.07 | 33.27 | 33.9 | 36,668,600 | 0.938067 | 0 | 0 |
2024-01-02 | BIL | 91.4 | 91.41 | 91.4 | 91.4 | 9,445,500 | 0.886031 | 0 | 0 |
2024-01-02 | BKNG | 140.8 | 141.2796 | 138.5888 | 139.2224 | 8,335,000 | 0.976841 | 0 | 0 |
2024-01-02 | BRK-B | 356.32 | 362.57 | 355.94 | 362.46 | 4,737,000 | 1 | 0 | 0 |
2024-01-02 | CAT | 293.43 | 296.64 | 291.35 | 292.71 | 2,433,100 | 0.961397 | 0 | 0 |
2024-01-02 | CDNS | 269.16 | 269.2 | 260.28 | 262.9 | 1,594,400 | 1 | 0 | 0 |
2024-01-02 | COST | 655.58 | 657.8 | 649.5 | 650.65 | 2,657,700 | 0.983234 | 0 | 0 |
2024-01-02 | CRM | 260.54 | 260.56 | 253.77 | 256.13 | 4,741,400 | 0.981267 | 0 | 0 |
2024-01-02 | CRWD | 62.8675 | 63.0725 | 60.7575 | 61.7225 | 15,027,200 | 1 | 0 | 0 |
2024-01-02 | CSCO | 50.49 | 50.86 | 50.25 | 50.51 | 20,242,900 | 0.926967 | 0 | 0 |
2024-01-02 | CVX | 149.99 | 151.48 | 149.24 | 149.48 | 8,879,700 | 0.891335 | 0 | 0 |
2024-01-02 | DE | 399.12 | 406.58 | 398 | 400.91 | 1,414,600 | 0.964226 | 0 | 0 |
2024-01-02 | DIA | 374.8 | 377.82 | 374.8 | 377.09 | 5,776,300 | 0.957469 | 0 | 0 |
2024-01-02 | DIS | 90.1 | 91.48 | 89.73 | 90.71 | 10,587,600 | 0.972723 | 0 | 0 |
2024-01-02 | DUK | 96.52 | 98.28 | 96.13 | 97.81 | 3,352,200 | 0.903582 | 0 | 0 |
2024-01-02 | GE | 101.3647 | 102.2745 | 99.9521 | 100.5188 | 5,189,425 | 0.983656 | 0 | 0 |
2024-01-02 | GLD | 191.44 | 191.89 | 190.5 | 190.72 | 6,025,600 | 1 | 0 | 0 |
2024-01-02 | GOOGL | 138.55 | 139.45 | 136.48 | 138.17 | 23,711,200 | 0.990574 | 0 | 0 |
2024-01-02 | GS | 383 | 389.14 | 382.19 | 388.3 | 2,263,100 | 0.943018 | 0 | 0 |
2024-01-02 | HD | 344.21 | 347.3 | 343.22 | 345.08 | 2,833,600 | 0.932112 | 0 | 0 |
2024-01-02 | HON | 206.452 | 208.1819 | 205.9183 | 206.5904 | 3,169,238 | 0.945391 | 0 | 0 |
2024-01-02 | HYG | 77.13 | 77.24 | 77.04 | 77.13 | 41,547,100 | 0.845674 | 0 | 0 |
2024-01-02 | IEF | 95.94 | 96.15 | 95.92 | 96.03 | 9,444,000 | 0.898929 | 0 | 0 |
2024-01-02 | INTU | 622.25 | 623.06 | 599.63 | 603.76 | 1,666,700 | 0.973804 | 0 | 0 |
2024-01-02 | ISRG | 333 | 336.2 | 328.38 | 330.98 | 1,379,200 | 1 | 0 | 0 |
2024-01-02 | IWM | 199.4 | 201.62 | 198.44 | 199.52 | 44,341,500 | 0.970796 | 0 | 0 |
2024-01-02 | JNJ | 156.93 | 161.64 | 156.79 | 159.97 | 9,622,100 | 0.925634 | 0 | 0 |
2024-01-02 | JPM | 169.09 | 172.17 | 168.91 | 172.08 | 9,977,400 | 0.938359 | 0 | 0 |
2024-01-02 | KO | 58.8 | 59.88 | 58.67 | 59.82 | 16,322,600 | 0.924596 | 0 | 0 |
2024-01-02 | LIN | 409.78 | 411.73 | 406.7 | 409.58 | 1,571,900 | 0.965319 | 0 | 0 |
2024-01-02 | LLY | 580.41 | 592.72 | 579.05 | 592.2 | 3,226,700 | 0.981419 | 0 | 0 |
2024-01-02 | LOW | 220.25 | 222.6 | 218.46 | 218.65 | 3,277,800 | 0.941356 | 0 | 0 |
2024-01-02 | LQD | 109.94 | 110.17 | 109.79 | 109.91 | 20,622,000 | 0.87995 | 0 | 0 |
2024-01-02 | MA | 424.09 | 424.28 | 419.56 | 421.89 | 2,463,700 | 0.982255 | 0 | 0 |
2024-01-02 | MCD | 295.05 | 297.28 | 295.05 | 297.04 | 4,458,400 | 0.935744 | 0 | 0 |
2024-01-02 | META | 351.32 | 353.16 | 340.01 | 346.29 | 19,042,200 | 0.990513 | 0 | 0 |
2024-01-02 | MRK | 109.16 | 113.32 | 109.16 | 113.24 | 11,920,100 | 0.919189 | 0 | 0 |
2024-01-02 | MSFT | 373.86 | 375.9 | 366.77 | 370.87 | 25,258,600 | 0.979098 | 0 | 0 |
2024-01-02 | NEE | 60.55 | 62.15 | 60.18 | 61.57 | 10,064,300 | 0.922189 | 0 | 0 |
2024-01-02 | NFLX | 48.319 | 48.465 | 46.186 | 46.85 | 50,494,000 | 1 | 0 | 0 |
2024-01-02 | NKE | 107.61 | 107.85 | 106.45 | 106.55 | 9,326,600 | 0.935846 | 0 | 0 |
2024-01-02 | NOW | 139.41 | 139.872 | 135.086 | 137.504 | 5,654,500 | 1 | 0 | 0 |
2024-01-02 | NVDA | 49.244 | 49.295 | 47.595 | 48.168 | 411,254,000 | 0.99711 | 0 | 0 |
2024-01-02 | O | 57.11 | 58.58 | 57.01 | 58.57 | 5,815,300 | 0.854984 | 0 | 0 |
2024-01-02 | ORCL | 104.52 | 104.68 | 102.44 | 104.06 | 9,597,500 | 0.965274 | 0 | 0 |
2024-01-02 | PANW | 145.97 | 146.06 | 142.415 | 144.46 | 5,767,800 | 1 | 0 | 0 |
2024-01-02 | PEP | 169.52 | 172.99 | 169.19 | 172.91 | 5,743,100 | 0.904212 | 0 | 0 |
2024-01-02 | PG | 146.36 | 149.41 | 146.31 | 148.74 | 7,238,400 | 0.923781 | 0 | 0 |
2024-01-02 | PLD | 132.84 | 134.71 | 132.1 | 134.62 | 2,899,100 | 0.913748 | 0 | 0 |
2024-01-02 | QCOM | 142.19 | 142.2 | 138.78 | 140.23 | 8,495,800 | 0.944424 | 0 | 0 |
2024-01-02 | QQQ | 405.84 | 406.09 | 400.24 | 402.59 | 58,026,900 | 0.985719 | 0 | 0 |
2024-01-02 | RSP | 157.02 | 158.35 | 156.77 | 157.85 | 9,816,500 | 0.956321 | 0 | 0 |
2024-01-02 | RTX | 84.05 | 86.02 | 83.96 | 85.22 | 10,455,600 | 0.948883 | 0 | 0 |
2024-01-02 | SBUX | 95.45 | 95.95 | 93.62 | 93.67 | 8,859,700 | 0.932439 | 0 | 0 |
2024-01-02 | SGOV | 100.29 | 100.3 | 100.28 | 100.29 | 5,286,200 | 0.888213 | 0 | 0 |
2024-01-02 | SHOP | 76.44 | 76.63 | 72.91 | 73.83 | 13,134,800 | 1 | 0 | 0 |
2024-01-02 | SHV | 110.13 | 110.13 | 110.12 | 110.12 | 2,636,900 | 0.885988 | 0 | 0 |
2024-01-02 | SHY | 81.94 | 81.97 | 81.93 | 81.94 | 4,983,500 | 0.898625 | 0 | 0 |
2024-01-02 | SNPS | 507.95 | 509.5 | 496.81 | 498.97 | 1,277,700 | 1 | 0 | 0 |
2024-01-02 | SO | 69.67 | 70.96 | 69.6 | 70.85 | 3,955,100 | 0.910876 | 0 | 0 |
2024-01-02 | SPGI | 415.5818 | 416.7455 | 410.369 | 412.8571 | 1,142,194 | 0.978228 | 0 | 0 |
2024-01-02 | SPY | 472.16 | 473.67 | 470.49 | 472.65 | 123,623,700 | 0.968312 | 0 | 0 |
2024-01-02 | TLT | 98.23 | 98.65 | 98 | 98.31 | 47,797,400 | 0.884896 | 0 | 0 |
2024-01-02 | TMO | 529.56 | 546.1 | 527.37 | 544.15 | 2,310,800 | 0.991037 | 0 | 0 |
2024-01-02 | TMUS | 160.33 | 163 | 160.33 | 161.91 | 5,049,700 | 0.954957 | 0 | 0 |
2024-01-02 | TSLA | 250.08 | 251.25 | 244.41 | 248.42 | 104,654,200 | 1 | 0 | 0 |
2024-01-02 | TXN | 168.85 | 169.93 | 167.41 | 169.26 | 5,652,500 | 0.921646 | 0 | 0 |
2024-01-02 | UNH | 526.84 | 539.66 | 526.67 | 539.34 | 3,415,700 | 0.94229 | 0 | 0 |
2024-01-02 | UNP | 244.98 | 245.99 | 242.75 | 243.43 | 2,622,200 | 0.941295 | 0 | 0 |
2024-01-02 | USO | 67.56 | 67.56 | 65.48 | 65.77 | 7,908,600 | 1 | 0 | 0 |
2024-01-02 | UUP | 27.29 | 27.33 | 27.26 | 27.32 | 2,109,400 | 0.925221 | 0 | 0 |
2024-01-02 | V | 259.61 | 260.09 | 257.67 | 258.87 | 5,471,000 | 0.977723 | 0 | 0 |
2024-01-02 | VTI | 235.65 | 236.51 | 234.77 | 235.63 | 4,847,600 | 0.966576 | 0 | 0 |
2024-01-02 | WMT | 52.4167 | 53.18 | 52.3967 | 53.0967 | 23,539,800 | 0.972593 | 0 | 0 |
2024-01-02 | XLB | 42.46 | 42.765 | 42.4 | 42.695 | 19,924,800 | 0.950779 | 0 | 0 |
2024-01-02 | XLC | 72.31 | 72.42 | 71.665 | 72.25 | 6,003,800 | 0.968373 | 0 | 0 |
2024-01-02 | XLE | 42.215 | 42.76 | 42.155 | 42.37 | 39,301,800 | 0.918414 | 0 | 0 |
2024-01-02 | XLF | 37.5 | 37.78 | 37.43 | 37.76 | 31,351,900 | 0.959287 | 0 | 0 |
2024-01-02 | XLI | 113.23 | 113.94 | 112.49 | 112.84 | 8,558,700 | 0.964282 | 0 | 0 |
2024-01-02 | XLK | 94.93 | 95.08 | 93.125 | 93.72 | 23,545,400 | 0.983492 | 0 | 0 |
2024-01-02 | XLP | 71.77 | 72.96 | 71.77 | 72.84 | 12,912,000 | 0.928906 | 0 | 0 |
2024-01-02 | XLRE | 39.97 | 40.45 | 39.78 | 40.43 | 6,591,500 | 0.911887 | 0 | 0 |
2024-01-02 | XLU | 31.53 | 32.2 | 31.465 | 32.125 | 30,758,600 | 0.923038 | 0 | 0 |
2024-01-02 | XLV | 136.04 | 139.05 | 135.97 | 138.78 | 11,785,900 | 0.955506 | 0 | 0 |
2024-01-02 | XLY | 89.125 | 89.165 | 88.235 | 88.575 | 15,369,600 | 0.977173 | 0 | 0 |
2024-01-02 | XOM | 100.92 | 103.1 | 100.85 | 102.36 | 23,483,000 | 0.914435 | 0 | 0 |
2024-01-02 | ^VIX | 13.22 | 14.23 | 13.1 | 13.2 | 0 | 1 | 0 | 0 |
2024-01-02 | ^VIX3M | 15.89 | 16.17 | 15.45 | 15.48 | 0 | 1 | 0 | 0 |
2024-01-02 | ^VIX9D | 13.71 | 13.93 | 12.55 | 12.59 | 0 | 1 | 0 | 0 |
2024-01-02 | ^VVIX | 91.86 | 92.82 | 87.59 | 88.09 | 0 | 1 | 0 | 0 |
2024-01-03 | AAPL | 184.22 | 185.88 | 183.43 | 184.25 | 58,414,500 | 0.987955 | 0 | 0 |
2024-01-03 | ABBV | 160.59 | 161.75 | 159.84 | 160.46 | 5,301,400 | 0.905054 | 0 | 0 |
2024-01-03 | ADBE | 574.58 | 577.3 | 570.19 | 571.79 | 2,478,000 | 1 | 0 | 0 |
2024-01-03 | AMD | 135.71 | 137.43 | 133.74 | 135.32 | 61,988,600 | 1 | 0 | 0 |
rta-bench: a benchmark for LLM trading agents, with news replayed by the minute it was published
Most news backtests join headlines to prices by calendar date, so a headline published at 16:05 is traded at the 16:00 close that came before it. rta-bench-1 is a public benchmark and leaderboard for LLM trading agents in which every headline arrives at the minute it was published and every order fills at the next open, through three frozen markets: the 2020 crash, the 2022 bear market and 2026 to date.
This dataset holds the frozen daily prices of the three markets (with the VIX term structure) and their manifests. The leaderboard's scores are on the website and in the Space. The benchmark replays licensed news headlines, which are not distributed with the data; this card gives the method, the numbers and the conclusions.
🏆 Leaderboard · GitHub · How the benchmark works · Benchmarks compared
| Period (split) | Test window | What happened | Headlines the benchmark replays (not redistributed) | Price rows here (with two warm-up years) |
|---|---|---|---|---|
2026 |
2026-01-02 to 2026-11-11, 217 trading days | A rising market: SPY +12.1%, with a -8.9% drawdown in March | 16,945 (headlines and SEC 8-K filings) | 69,738 |
2022 |
2022-01-03 to 2022-12-30, 251 trading days | The rate-hike bear market: SPY -18.2%, the Nasdaq 100 -32.6% | 29,300 (headlines) | 72,466 |
2020 |
2020-01-02 to 2020-12-31, 253 trading days | The COVID crash and rebound: SPY -34% in 23 trading days, VIX at 82.7, +18.3% for the year | 19,618 (headlines) | 71,490 |
Why by the minute
A calendar-date join is the most common way a news backtest sees the future. Take one simple headline rule on the 2026 data: +1 when a company beat estimates or raised its outlook, -1 when it missed or cut, held for one day.
| How news is joined to prices | Trades | Mean per trade |
|---|---|---|
| Calendar date (close to next close) | 444 | +32.0 bp |
| Publication time (next open to the open after) | 458 | -6.7 bp |
All of the apparent edge comes from the 178 headlines published after the close, which make +103 bp per trade under the calendar join because the trade buys at a price set before the news came out. The same leak shows up in daily news factors across all three periods. Lexicon sentiment has a one-day rank IC of 0.032 (t = 3.4) when joined by calendar date and tested close to close, and 0.005 (t = 0.6) when joined by publication time and traded from the next open. For headline sentiment scored by gpt-5.6-luna, the calendar join gives 0.050 (t = 5.6) against 0.021 (t = 2.45). Sources: the news backtesting guide and the news factor study.
The benchmark: rta-bench-1
Every model trades the same 56 large-cap US stocks (11 sectors) through each period. It decides after every fifth close, on prices, indicators, its portfolio and up to three headlines per stock published before that close. Orders fill at the next trading day's open with realistic costs: 5 bps, a tenth of the overnight gap and square-root market impact, with a volume cap. Long only, at most 10 positions, $100,000 to start; uninvested cash earns T-bills.
- Six tasks:
2026,2022and2020, each in two forms, with seeds 0, 1 and 2 per task.- Blind: tickers become S01, S02, ...; dates become day numbers; prices are rebased to 100; names, people, products, dates and dollar amounts inside headlines are masked.
- Named: everything is shown.
- Scores split at each model's training cutoff.
clean: the window starts after the cutoff.partly: the cutoff falls inside the window; the run is also scored from the cutoff on.may_remember: the window ends before the cutoff; listed apart and never ranked as skill.- The named-minus-blind gap and a separate memory probe say how much a model remembers.
- A result is a file.
- Each run is recorded order by order, with the SHA-256 of every observation the model was shown, so its trades can be replayed through the backtest without calling the model.
- A look-ahead check replays the orders on copies of the market whose prices and headlines after three cut days are random: the observations up to each cut must be byte for byte the same.
The board on 2026-11-13 (26 entries, 4 models; repository commit 28d9734)
The board is filled as runs finish; the live version, with seeds, factor attributions and decision statistics, is on the website. Models so far: claude-haiku-5.5, claude-sonnet-4.6, gemini-3.5-flash, gpt-5.6-luna.
2026, blind, scored only on the days after each model's training cutoff (mean over seeds, range in brackets):
| Model | Training cutoff | Seeds | Scored from | Return | SPY, same days | Equal weight, same days | Sharpe |
|---|---|---|---|---|---|---|---|
| claude-sonnet-4.6 | 2026-01 | 1 | 2026-01-30 | +6.2% | +10.7% | +10.6% | 0.34 |
| gemini-3.5-flash | 2025-01 | 1 | 2026-01-02 | +5.4% | +12.1% | +11.4% | 0.25 |
| gpt-5.6-luna | 2026-02-16 | 3 | 2026-02-13 | +1.7% (-1.9% to +5.3%) | +12.3% | +10.7% | -0.15 |
| claude-haiku-5.5 | 2026-06 | 1 | 2026-06-30 | -0.6% | +2.0% | +5.7% | -0.99 |
Every period, blind, whole window. 2020 and 2022 are before every listed model's cutoff (may remember), so these rows describe behaviour, not skill:
| Model | Period | Label | Seeds | Return | Sharpe | Max drawdown | vs equal weight |
|---|---|---|---|---|---|---|---|
| claude-haiku-5.5 | 2026 | partly | 1 | +2.4% | -0.06 | -5.6% | -9.0% |
| claude-sonnet-4.6 | 2026 | partly | 1 | +8.1% | 0.44 | -9.0% | -3.3% |
| gemini-3.5-flash | 2026 | clean | 1 | +5.4% | 0.25 | -9.6% | -6.0% |
| gpt-5.6-luna | 2026 | partly | 3 | +5.4% | 0.42 | -4.8% | -6.0% |
| claude-haiku-5.5 | 2022 | may remember | 1 | +1.4% | -0.03 | -0.7% | +13.7% |
| claude-sonnet-4.6 | 2022 | may remember | 1 | +0.4% | -0.03 | -13.5% | +12.7% |
| gemini-3.5-flash | 2022 | may remember | 1 | -1.2% | -0.24 | -8.7% | +11.2% |
| gpt-5.6-luna | 2022 | may remember | 3 | -2.5% | -0.50 | -8.5% | +9.9% |
| claude-haiku-5.5 | 2020 | may remember | 1 | +7.4% | 0.77 | -9.0% | -32.7% |
| claude-sonnet-4.6 | 2020 | may remember | 1 | +25.4% | 1.38 | -14.0% | -14.7% |
| gemini-3.5-flash | 2020 | may remember | 1 | +5.6% | 0.35 | -19.7% | -34.4% |
| gpt-5.6-luna | 2020 | may remember | 3 | +2.3% | 0.20 | -15.3% | -37.8% |
| Equal weight, all 56 | 2026 | +11.4% | 0.85 | -7.2% | — | ||
| SPY buy and hold | 2026 | +12.1% | 0.77 | -8.9% | — | ||
| Trend following (rule) | 2026 | +19.0% | 1.27 | -5.0% | +7.5% | ||
| Equal weight, all 56 | 2022 | -12.4% | -0.53 | -23.1% | — | ||
| SPY buy and hold | 2022 | -18.6% | -0.80 | -24.5% | — | ||
| Trend following (rule) | 2022 | -5.9% | -0.47 | -17.8% | +6.4% | ||
| Equal weight, all 56 | 2020 | +40.1% | 1.13 | -32.5% | — | ||
| SPY buy and hold | 2020 | +17.2% | 0.63 | -33.7% | — | ||
| Trend following (rule) | 2020 | +67.4% | 2.18 | -13.6% | +27.4% |
The configs
prices (default): splits 2026, 2022, 2020
| Column | Type | Meaning |
|---|---|---|
date |
date | Trading day (New York) |
symbol |
string | Yahoo Finance ticker, for example AAPL, BRK-B, ^VIX |
open, high, low, close |
float | Prices as traded, adjusted for splits as of the period's last day |
volume |
int | Volume of the day; 0 for the volatility indices |
adj |
float | Adj Close / Close: multiply the four prices by it to get total-return prices that include dividends |
dividend |
float | Cash dividend paid that day, else 0 |
split |
float | Split ratio that day, else 0 |
Each split is a test window plus the two years before it, which warm up slow indicators such as EMA200 and the 52-week high. The 2026 split starts on 2024-01-02, 2022 on 2020-01-02 and 2020 on 2018-01-02. The splits overlap in time (2020 is in both 2020 and 2022), and each keeps its own split and dividend adjustment as of its last day. Treat every split as its own frozen market rather than joining them.
Symbols:
- 64 stocks, the package's watchlists: AAPL, ABBV, ADBE, AMD, AMZN, AVGO, BAC, BKNG, BRK-B, CAT, CDNS, COST, CRM, CRWD, CSCO, CVX, DE, DIS, DUK, GE, GOOGL, GS, HD, HON, INTU, ISRG, JNJ, JPM, KO, LIN, LLY, LOW, MA, MCD, META, MRK, MSFT, NEE, NFLX, NKE, NOW, NVDA, O, ORCL, PANW, PEP, PG, PLD, QCOM, RTX, SBUX, SHOP, SNPS, SO, SPGI, TMO, TMUS, TSLA, TXN, UNH, UNP, V, WMT, XOM. CrowdStrike starts on 2019-06-12.
- 25 ETFs: SPY, QQQ, DIA, IWM, RSP, VTI, the eleven sector funds, and BIL, SGOV, SHV, SHY, IEF, TLT, LQD, HYG.
- Volatility indices:
^VIX9D,^VIX,^VIX3Mand^VVIX. - Cross-asset ETFs: GLD, USO and UUP, plus IBIT in 2026.
The benchmark trades the 56 stocks of the package's megacap list.
How the benchmark uses news
The news is not in this dataset; this is how the benchmark treats it.
- Sources. Headlines (titles, times and tickers) from a licensed news feed for all three periods, plus SEC 8-K filings with the time SEC accepted them for 2026. The headline archive is licensed, so only results computed from it are published.
- Replay rule. A decision after day t's close (16:00 New York, 13:00 on early-close days) sees only items published strictly before that close. An item stamped at 16:00 sharp, when many earnings releases go out, is first visible on day t+1.
- Tying a headline to a stock. In 2026 a headline belongs to a stock when the feed tagged the stock and the headline names the company or its ticker. The 2020 and 2022 exports had lost their ticker tags, so there a headline belongs to every stock whose company name or ticker it contains.
- Your own news. The Python package replays any timestamped archive you supply with the same rules: SEC 8-K filings, which are free, your own news API key, or a CSV. See backtesting with news.
Load it
from datasets import load_dataset
prices = load_dataset("jacksonwangpro1/rta-bench-llm-trading-benchmark", "prices", split="2022") # or "2026", "2020"
pandas (needs huggingface_hub for the hf:// paths):
import pandas as pd
base = "hf://datasets/jacksonwangpro1/rta-bench-llm-trading-benchmark/data/"
prices = pd.read_parquet(base + "prices_2022.parquet") # 2026: prices.parquet
prices["close_tr"] = prices["close"] * prices["adj"] # total-return close, dividends included
Check the files
The raw/ folder holds the package's price files byte for byte, with one manifest per period (manifest.json for 2026, manifest_2022.json, manifest_2020.json) that records the data version and the SHA-256 of each file. The Parquet files hold the same rows and columns, with date stored as a date.
| Period | Data version | File | SHA-256 |
|---|---|---|---|
| 2026 | 2026-11-11 |
market_2026.csv.gz |
35183f30262ad64966bd3b5fb75b908c9e604c394c85faf8be9527d034491b43 |
| 2022 | 2022-12-30 |
market_2022.csv.gz |
6377384cd1902329bf69dcb85338335b15f0e9e4485a753ad005cd370c9238fc |
| 2020 | 2020-12-31 |
market_2020.csv.gz |
5b151a2375efe4f0ce2da966e3976c67504f063df717b162f716fc484801a2e1 |
shasum -a 256 raw/*.gz
When you report results, give the period, its data version and the SHA-256 of the files you used.
Sources
- Prices: Yahoo Finance daily bars, downloaded with yfinance (
auto_adjust=False, actions=True).scripts/build_market_bundle.py --period 2022in the GitHub repository rebuilds a period. - Headlines used by the benchmark (not included): a licensed headline feed with publication times, and SEC EDGAR 8-K filings for 2026.
Limits
- Survivorship. The 64 stocks are today's large caps, chosen with hindsight, and the hindsight grows with distance: buying and holding the 56 benchmark stocks returned +40.1% in 2020 against SPY's +17.2%. Compare models with each other and with the rule anchors, not only with SPY.
- One headline feed. The benchmark's news is one licensed headline feed plus, in 2026, SEC filings, so most stories have a single source.
- Memory. Language models trained on text from these years may remember how the stocks did. Blind mode lowers the chance of recall but cannot rule it out, which is why the benchmark splits scores at each model's training cutoff.
- Short windows. Ten to twelve months per period; the Sharpe interval in each entry says how short. Read differences smaller than the range over seeds as noise.
- No fundamentals or analyst data: no free source has them point-in-time.
License and terms
The code of Risk-Aware Trading Agents is MIT-licensed. The prices here stay under the terms of Yahoo Finance, which is why the license is other. The data is shared for research and education, as is, without warranty. Details are in TERMS.md. If you hold rights to any of it and want something removed, please open an issue on GitHub. Nothing here is investment advice.
Citation
Until the benchmark paper is out, please cite the GitHub repository. Once the paper is published, cite the paper instead; the repository's README will give the reference.
@software{riskaware_rta_bench,
title = {Risk-Aware Trading Agents: a benchmark for LLM trading agents with news replayed by the minute it was published},
author = {{Risk-Aware Trading Agents contributors}},
year = {2026},
url = {https://github.com/jackson-cmd/risk-aware-trading-agents},
note = {rta-bench-1 prices: periods 2026 (version 2026-11-11), 2022 (2022-12-30) and 2020 (2020-12-31), https://huggingface.co/datasets/jacksonwangpro1/rta-bench-llm-trading-benchmark}
}
- Downloads last month
- 126