Research benchmark only. Not financial advice. Simulated capital — crypto trading can lose all capital.
DeepSeek Crypto Trading: Headline Record vs Settled Book
DeepSeek's headline record and its settled book disagree. On the headline it went 2 of 5 profitable across Seasons 3-6 and Season 8, best +11.85% and worst -11.66%. On the settled book only Season 4 closed green: Season 5's +11.85% sat on open positions over a -2.26% realized book, and Season 3, Season 6 and Season 8 all booked losses.
- 5 completed seasons
- 2 profitable
- +11.85% best season (Season 5)
- -11.66% worst season (Season 3)
Season-by-season record
Completed TradeRank Arena seasons only. Win rate includes open positions at season close. See the live dashboard →
| Season | Return | Rank | Trades | Win % | Max DD | Realized | Unrealized |
|---|---|---|---|---|---|---|---|
| Season 3 | -11.66% | 8 / 9 | 35 | 23% | +12.98% | -11.22% | -0.45% |
| Season 4 | +5.40% | 2 / 9 | 13 | 31% | +4.54% | +2.59% | +2.81% |
| Season 5 | +11.85% | 2 / 10 | 8 | 75% | +9.76% | -2.26% | +14.11% |
| Season 6 | -3.17% | 10 / 11 | 15 | 47% | +9.72% | -2.85% | -0.32% |
| Season 8 | -1.89% | 13 / 14 | 7 | 14% | +3.58% | -1.75% | -0.15% |
Performance at a glance
Version history
| Season | Model version | Notes |
|---|---|---|
| Season 3 | DeepSeek V3.2 | — |
| Season 4 | DeepSeek V4 Pro | — |
| Season 5 | DeepSeek V4 Pro | — |
| Season 6 | DeepSeek V4 Pro | — |
| Season 8 | DeepSeek V4 Pro 0813 | — |
Representative decisions
Outcomes are reconstructed from position state changes between consecutive daily snapshots — not trade fills.
- Gain — open_short ADA (Season 3): Weekly ▼, Daily ▼, 4h ▼, composite 70. Bearish momentum. Positive funding.
- Loss — open_short UNI (Season 3): Weekly ▼, Daily ▼, 4h ▼, composite 70. Downtrend intact. Positive funding.
In its own words
Weekly ▼, Daily ▼, 4h ▼, composite 70. Bearish momentum. Positive funding.
Weekly ▼, Daily ▼, 4h ▼, composite 70. Downtrend intact. Positive funding.
Can DeepSeek trade crypto?
DeepSeek does not place trades on its own. In TradeRank Arena it makes autonomous daily decisions on simulated capital under identical market data, prompt rules, and risk controls, and every decision and trade is logged publicly. The record above is that controlled benchmark — not a live brokerage account, and not financial advice.
Methodology
- Season 3 (2026-03-23 to 2026-04-26): 37 crypto assets, daily (24h cycle), $10,000 starting capital, 0.1% fee.
- Season 4 (2026-04-26 to 2026-05-23): 7 crypto assets, daily (24h cycle), $10,000 starting capital, 0.1% fee.
- Season 5 (2026-05-23 to 2026-06-20): 10 crypto assets, daily (24h cycle), $10,000 starting capital, 0.1% fee.
- Season 6 (2026-06-20 to 2026-07-18): 10 crypto assets, daily (24h cycle), $10,000 starting capital, 0.1% fee.
- Season 8 (2026-08-15 to 2026-09-12): 10 crypto + 50 equity assets, daily (24h cycle), $10,000 starting capital, 0.1% fee.
Limitations
- returnPct includes unrealized P&L; realized and unrealized are reported separately per season (a season leader has previously carried negative realized P&L while leading on unrealized gains).
- Win rate is taken from report standings, which count still-open positions as trades; it is not a realized closed-trade win rate and should be read alongside the realized/unrealized split.
- Representative-decision outcomes are position state change between consecutive daily equity snapshots — not a trade fill. Same-cycle round-trips are irreducible and season-end opens are mark-to-market only.
- Cross-season only: prompts, model versions, cadence, asset universe and market regimes changed between seasons — this is not one controlled experiment.
- Capital was simulated, prices were live, modeled fees applied; execution omitted slippage, market impact, borrow costs and real capital risk.
- Sample is short (5 completed seasons); behavior is described with sample size, never as a permanent trait of DeepSeek.
- Hold-time is not reported: no hold-time field exists in standings and the "0.0 hours" text in report prose is a placeholder, not a measurement. Profit factor is not archived and is omitted.
- One or more seasons were force-liquidated at close (every position sold, fee-bearing), so those seasons closed under different conditions from seasons that retained open exposure. Every return on this page — including the cross-season aggregates — is the return the standings finalized. The last pre-liquidation daily mark is recorded in the data as comparabilityReturnPct for reference and is not used as a result here.
Frequently asked questions
Can DeepSeek trade crypto?
Yes, as an autonomous agent in a research benchmark. On TradeRank a DeepSeek model made daily trading decisions at live prices, with a $10,000 simulated stake and a modeled 0.1% fee, across the 5 completed seasons this record covers, Seasons 3-6 and Season 8. It finished 2 of the 5 in profit. This is paper trading in a research benchmark and is not investment advice.
How has DeepSeek performed in TradeRank's crypto trading competition?
-11.66% and 8th of 9 in Season 3, +5.40% and 2nd of 9 in Season 4, +11.85% and 2nd of 10 in Season 5, -3.17% and 10th of 11 in Season 6, and -1.89% and 13th of 14 in Season 8: 2 of the 5 profitable on the headline. On the realized book only Season 4 closed positive.
Is DeepSeek's crypto trading record all one model version?
No. The seat ran DeepSeek V3.2 in Season 3, DeepSeek V4 Pro in Seasons 4 through 6 and DeepSeek V4 Pro 0813 in Season 8. The asset list also moved, from 37 crypto names to 7, then 10, and US stocks joined partway through Season 6, so any result belongs to a build and a season together.
Was DeepSeek's best crypto trading season a real profit?
On the settled book it was a loss: Season 5's realized P&L closed at -2.26%, and the +11.85% came from +14.11% of open marks at the close. The +11.85% is still the official simulated result and placed 2nd of 10. The headline and the booked figure measure different things.
How does DeepSeek's crypto trading compare to Alpha Arena?
The arena, the build and the rules all differ. This record is 5 completed seasons of simulated trading by DeepSeek V3.2, V4 Pro and V4 Pro 0813. nof1's Alpha Arena pioneered the format, and in its Season 1, with real capital on Hyperliquid in late 2025, a DeepSeek V3.1 build finished runner-up behind Qwen 3 Max. The 2 results sit side by side and do not add up to one record.
What can and can't this DeepSeek record establish?
It states what happened. DeepSeek went 2 of 5 profitable, best +11.85% and worst -11.66%, with only Season 4's realized book closing positive. It cannot establish a lasting trait: 5 seasons across 3 builds and changing markets is too small a sample.