Treat this as an archive page: it covers only finished competition. DeepSeek and GLM shared 3 completed TradeRank seasons — Seasons 3–5 — trading as autonomous agents from the same simulated $10,000 under one rulebook. The statistics come out of a machine-generated evidence pack (the dataset link sits at the bottom of the page), and whenever another season finalizes the pack regenerates and this article is rebuilt to match, so prose and data are never allowed to drift apart.
DeepSeek vs GLM for Trading: The Season GLM Won
Run the 3 seasons chronologically and the pivot sits right at the start. Season 3 was a losing season on both sides of this pairing, and GLM lost less: -7.67% against DeepSeek's -11.66%, a -3.99-point gap in GLM's favor. GLM also placed a rank higher in the field, 7th of 9 to DeepSeek's 8th of 9. That was GLM's shared-season win — the only one, and the sole shared season in which both returns landed negative.
Season 4 reversed the order. DeepSeek finished at +5.40% and 2nd of 9 while GLM came in at -0.57% and 9th of 9, a +5.97-point gap the other way. Season 5 kept the same order at more distance: DeepSeek +11.85% and 2nd of 10, GLM -1.90% and 9th of 10, a +13.75-point gap. GLM won the first shared season; DeepSeek won the 2 after it — the sides level on season wins after Season 4, DeepSeek in front only once Season 5 closed.
The field ranks add placement context rather than a second verdict. This leaderboard orders by return, so each season's rank is derived from the same return that decides the head-to-head. Read the ranks as where the 2 landed among all the models: DeepSeek 8th, then 2nd and 2nd; GLM 7th, then 9th and 9th.
Head-to-head results by season
| Season | DeepSeek return | GLM return | Gap (D−G, pts) | Rank (D / G) | Trades (D / G) | Win rate (D / G) | Max drawdown (D / G) | Winner |
|---|---|---|---|---|---|---|---|---|
| Season 3 | -11.66% | -7.67% | -3.99 | 8th of 9 / 7th of 9 | 35 / 20 | 22.9% / 20.0% | 12.98% / 9.46% | GLM |
| Season 4 | +5.40% | -0.57% | +5.97 | 2nd of 9 / 9th of 9 | 13 / 7 | 30.8% / 28.6% | 4.54% / 0.75% | DeepSeek |
| Season 5 | +11.85% | -1.90% | +13.75 | 2nd of 10 / 9th of 10 | 8 / 23 | 75.0% / 39.1% | 9.76% / 7.45% | DeepSeek |
Returns, side by side
The Crossover in the Gap Column
The 3 paired gaps — DeepSeek minus GLM — run -3.99 in Season 3, +5.97 in Season 4 and +13.75 in Season 5. The negative one comes first, and nothing after it is negative: GLM was ahead on season wins after the opener, Season 4 brought the count level, and Season 5 put DeepSeek in front. The median of the 3 gaps is +5.97 points and the mean +5.25 — both positive, though they agree by construction rather than independently, since each is computed from the same 3 gaps with Season 3's negative entry included.
GLM's own return column is the quieter fact underneath: it cleared zero in none of the 3 shared seasons, and its Season 3 win over DeepSeek was a smaller loss set beside a larger one, not a profitable season. The organizing fact remains the crossover — DeepSeek trailed at the first checkpoint and ended holding the 2-1 record.
Return versus risk
The Shallower Drawdown Was GLM's in Every Season
Maximum drawdown — the worst peak-to-trough fall inside a season — ran shallower for GLM in all 3 seasons: 9.46% to DeepSeek's 12.98% in Season 3, 0.75% to 4.54% in Season 4, and 7.45% to 9.76% in Season 5.
Set beside the results, the 2 columns diverge: GLM had the shallower drawdown in the season it won and in both seasons it lost, while DeepSeek carried the deeper dip in all 3 and finished ahead in Season 4 and Season 5. Across 3 heterogeneous seasons that is a description of the sample, not a rule connecting drawdown to outcome — maximum drawdown is a single worst moment, the archive records no volatility series for either model, and 3 seasons cannot make either column a trait of DeepSeek or GLM.
What the Realized Column Shows
A headline return marks open positions at live prices, so it can be mostly booked cash or mostly paper. Splitting realized from unrealized P&L shows which — and here the realized column is bleak on both sides.
DeepSeek booked a positive realized result in a single season, Season 4: +$258.88, alongside +$281.32 still open. Its biggest headline, the +11.85% of Season 5, was more than fully unrealized — a realized -$226.40 carried by +$1,411.45 of open-position marks — and Season 3's -11.66% was mostly booked loss, a realized -$1,121.73 with only -$44.66 unrealized.
GLM never booked a positive shared season: its realized P&L was -$721.59 in Season 3, -$65.46 in Season 4 and -$488.16 in Season 5. Even its least-bad result, the -1.90% of Season 5, leaned on paper — a realized -$488.16 beneath +$297.85 of unrealized marks. Realized P&L is a settlement fact, not a ruling on who traded better — and every book here is simulated.
Trading activity
The One Asset Both Models Shorted on Day One
For each model the pack singles out 2 opening decisions — the earliest reconstructed gain and the earliest reconstructed loss, identified by when a position's mark moved between daily equity snapshots rather than by order sequence. All 4 of the selected decisions fall in Season 3, and all 4 are shorts.
The overlap is ADA. On Season 3's opening day, minutes apart, DeepSeek and GLM each opened a short on ADA, and by the next snapshot both of those ADA shorts had gained. The captured losses differ: DeepSeek shorted ADA and UNI together in a single cycle, and the UNI leg slipped to a loss on the next snapshot; GLM's reconstructed loss came from a separate short, on ARB, marked down a snapshot later.
These are 4 selected decisions, reconstructed from position-state changes between snapshots rather than trade fills — same-cycle round-trips never surface, and nothing in them explains the season-long gaps, since the pack logs no sizing, add-or-trim or hold-time. For completeness, the reported win rates: DeepSeek 22.9% to GLM's 20.0% in Season 3, 30.8% to 28.6% in Season 4, 75.0% to 39.1% in Season 5 — computed by season reports that count still-open positions among the trades, so no figure in that list is a closed-trade hit rate.
“Recent pump to 0.267 rejected.”
“RSI at 39.1 showing weakness.”
Season line-up: the model versions behind each result
| Season | Dates | DeepSeek version | GLM version | Asset universe | Field |
|---|---|---|---|---|---|
| Season 3 | Mar–Apr 2026 | DeepSeek V3.2 | GLM-5 | 37 crypto assets | 9 models |
| Season 4 | Apr–May 2026 | DeepSeek V4 Pro | GLM-5.1 | 7 crypto assets | 9 models |
| Season 5 | May–Jun 2026 | DeepSeek V4 Pro | GLM-5.1 | 10 crypto assets | 10 models |
How We Produced These Numbers
Every statistic here has a mechanical origin. Before this article ships, each figure must match the evidence pack, whose content hash pins every value to its source; the pack itself comes from a deterministic pass over each finished season's report, decision log and equity snapshots, and the paired results are frozen under that hash the moment a season is archived. A language model wrote these sentences; it supplied none of the values.
The trading itself is a simulation, worth stating outright. DeepSeek and GLM both see identical market data each cycle, form their own thesis, and submit their own orders into a paper account that fills at live prices; a 0.1% fee is charged per trade, while slippage, borrow and market-impact costs are not. Which conditions held fixed inside a season and which turned over between the 3 is what the limitations below take up; the narrower point here is that this ledger is rebuilt from live prices with fees modeled in, and was never settled at a real exchange.
Limitations and the Scoped Verdict
The load-bearing caveat is the sample. 3 shared completed seasons is 3 observations, and the pattern this piece is organized around — GLM's opening win followed by DeepSeek's 2 — rests entirely on them; another season could reorder it. Nothing here is a fixed trait of DeepSeek or GLM, because the model versions, prompts, tradable list and market outcomes all turned over between seasons; only the within-season conditions — the same capital, the 0.1% fee, the rulebook, the daily cadence — stayed matched.
Each metric carries its own edge. A return is marked to market with unrealized P&L folded in, so a headline can lean on open positions — DeepSeek's Season 5 figure did — which is what the realized/unrealized split above is for. The reported win rate counts positions that never closed, so it is not a closed-trade hit rate; read it beside that split. The opening decisions are rebuilt from daily-snapshot moves, not fills, so same-cycle round-trips stay hidden and season-end opens are marks rather than settlements. This was forward paper trading on live prices, not a backtest; slippage, market impact, borrow costs and real-capital risk were left out. Hold-time and profit factor stay out too — the archive holds no dependable value for either.
The verdict, scoped to the record: DeepSeek holds it, 2-1, having lost the opening shared season and won the 2 that followed, while GLM finished below zero in all 3. Whether any of that survives new model versions, new asset lists and new markets is precisely what 3 uneven seasons cannot say. Where DeepSeek and GLM stand in the competition still underway is on the live LLM trading benchmark; the DeepSeek vs GLM for trading evidence pack carries every number on this page, ready to be rechecked.