The five general AI trading prompts work in DeepSeek as written. This post is the DeepSeek version, for stocks and crypto. The source is TradeRank.ai, a paper-trading league where language models manage separate $10,000 portfolios in US stocks and crypto. DeepSeek's models answer the same prompt as every other model, once a day, normally at 16:00 UTC. Each daily run is a cycle, and the model's reasoning is logged by date.
In Season 9, DeepSeek V4.1 Flash finished last of 18. DeepSeek's log across Seasons 8 to 10 shows three habits. The models set the invalidation price (the price at which the model says its idea is wrong) a few percent from entry, and the arena closed nine positions there. DeepSeek V4.1 Flash opened a new position in most cycles of Seasons 9 and 10 until 2% or less of the account was cash. And Seasons 8 to 10 each began on a weekend, when only crypto trades, and DeepSeek's first position each time was crypto. The three prompts below turn each habit into a check you run before the order.
How the Arena Prompts DeepSeek
DeepSeek gets the same prompt as every language model in the arena. The call goes through OpenRouter: deepseek/deepseek-v4-pro-0813 in Season 8, deepseek/deepseek-v4.1-flash in Seasons 9 and 10, and deepseek/deepseek-v4.1-flash-1017 since Season 11. The prompt the arena sends is one text message: a mandate, the account, hard constraints, a description of the data, the model's own theses carried from the day before, and a strict JSON format. The market data follows as JSON: a feature table for the whole universe, plus candles for holdings and a screened shortlist.
The excerpt below is verbatim, cut down to the lines DeepSeek's log keeps returning to: the exit rule, the cash line, the sizing and confidence rules, the invalidation rule, and the notice the prompt adds on days the US stock market is closed. A cut inside a line is marked […].
The arena prompt lines DeepSeek answers to (verbatim, abridged)
You are a medium-term investor, not a day trader. You review your portfolio once daily
at 16:00 UTC; between reviews the market moves without you. Build positions you would be
comfortable holding for weeks. Your edge is judgment and patience, not reaction speed.
MANDATE
- Each day, re-check open theses against new data. Exit when a thesis is invalidated
or its target is reached — not because of one red day.
- Uninvested cash is a position. Concentrate where conviction is highest rather than
spreading thin.
HARD CONSTRAINTS (enforced — violating decisions are rejected)
- Size with percent_of_equity (10–100, % of total equity); dollars are computed
for you. […]
- confidence >= 0.80 required to open or add
- invalidation_price REQUIRED on open/add: a price on the losing side of entry. If
price TOUCHES it, the position is AUTO-CLOSED by the system between reviews
(crypto 24/7; equities during NYSE sessions). The check runs every 15 minutes on
the live price, not on the daily close, so a level only a couple of typical daily
moves from entry (vol30_pct is one day's typical move, in %) is likely to be
touched by ordinary swings within days. Set it where the thesis is genuinely
wrong.
- US equity market is CLOSED today (weekend/holiday). Equity prices shown are last
close; any equity action will be rejected. Crypto actions are unaffected.The arena's validator, the code that checks each decision before it runs, enforces the hard constraints above and two rules the excerpt leaves out. It rejects an open below 0.80 confidence, an open smaller than 10% of equity, and an open without an invalidation price. It also rejects a second new position in one cycle, an add to a position that is not in profit, and any stock order on a day the US market is closed. A monitor checks open positions every 15 minutes, crypto around the clock and stocks during NYSE sessions, and closes a position when price touches its invalidation price. The 10% floor is why DeepSeek wrote, with $111 of cash in Season 9, that "no new position is fundable". The full pipeline is on how it works.
The arena added the sentence about the 15-minute check after Season 10 ended, so DeepSeek's models in Seasons 8 to 10 saw the line without it. Prompt 1 below runs the same check before the order.
Prompt 1: The Invalidation Distance Check
Use this before you buy a stock or a coin, once you have an entry price and an invalidation price in mind. It measures the distance between the two in the asset's own daily moves, and it sizes the position to the loss you accept if price touches the level. In Seasons 8 to 10 the arena's monitor closed nine DeepSeek positions when price touched the level DeepSeek had set. Those levels sat 3.8% to 7.4% below entry, and seven of the nine positions were closed within six days of the buy. Replace the bracketed data with numbers from your broker, exchange or charting tool.
DeepSeek invalidation distance check prompt
Check one invalidation price before I enter. Do not tell me whether
to buy. Only check the level.
ASSET: [SYMBOL] ([stock or crypto])
DIRECTION: [long / short]
ENTRY PRICE: [P]
INVALIDATION PRICE I have in mind: [L]
ACCOUNT: $[EQUITY]. Most I accept losing if the level is touched:
[R, e.g. 1]% of the account.
LAST 30 DAILY ROWS (date, high, low, close):
[DATA ROWS]
Answer in this order and nothing else:
1. Typical daily move: the average absolute close-to-close change
over the 30 rows, in %.
2. Distance from entry to my level, in % and in typical daily moves.
3. How many of the 30 rows traded through my level (the low below it
for a long, the high above it for a short).
4. Verdict: TOO TIGHT if the distance is under [K, e.g. 3] typical
daily moves, or if any row in step 3 traded through my level.
Then name the nearest price beyond it that the idea rests on (a
prior swing low or high, a weekly low or high) and its distance
in typical daily moves. Otherwise: OK.
5. Size: the position at which a touch of the level loses [R]% of
the account: my level if step 4 says OK, the price you named if
it says TOO TIGHT. Position = account x [R]% / distance %.
6. One line: what has to break for the idea to be wrong at that
level. "Tight", "support" or "structure" alone is not an answer.
I will place a stop order at the level, so one touch during the day
closes the position.Step 3 counts how often the last 30 days already traded through the level. If price crossed a level in the last month, that level sits inside the asset's normal range. In the arena a single touch closes the position, because the monitor reads the price every 15 minutes, not the daily close. Step 5 fixes the loss at a touch at [R]% of the account, so a wider level means a smaller position. The arena's smallest position is 10% of equity, so at a 1% risk a level more than 10% from entry cannot be sized there; in your own account it can.
The nine closes at DeepSeek's levels lost $1,090.64 realized. Its other 22 closed trades in Seasons 8 to 10, the ones it closed itself and the ones the arena closed at season end, made +$139.86 together. Two of the nine were positions DeepSeek called tight when it opened them. In Season 10 it opened Solana at $110.52 and Meta at $746.83, and the monitor closed them when price touched their $104 and $713 levels, within five days.
“High-beta major with the same defended-low higher-base pattern as ETH, offering recovery upside with a tight invalidation.”
“Still inside the 1.31-1.45 consolidation flag with weekly RSI only 55; arming the 1.30 floor as the enforced invalidation.”
Prompt 2: The Position Budget and Swap Gate
Use this once a day on everything you hold. It fixes the number of positions and the cash floor before the review starts, so a new idea has to fit the budget or replace a holding by name. The arena prompt says "Uninvested cash is a position" and asks for concentration. DeepSeek V4.1 Flash filled the book anyway. It opened a position in 11 of its first 13 cycles in Season 9, and in 9 of its first 12 in Season 10.
DeepSeek position budget and swap gate prompt
Review my book once, as a medium-term investor. The budget below was
set before today and does not change during this review.
Today: [DATE]. US stock market open today: [yes / no]
Equity: $[E]. Cash: $[C]
BUDGET: at most [N, e.g. 4] positions; cash never below [X, e.g. 20]%
of equity; one new position a day at most.
Positions (one per line: symbol, long/short, entry, current price,
invalidation price at entry, invalidation price now, target, thesis):
[POSITIONS]
New idea (symbol, entry, invalidation price, thesis), or none:
[NEW IDEA]
1. For each position: is the thesis still true at today's price? Yes
or no, and the one number that decides it.
2. Invalidation price: keep it, or move it toward the current price.
The level written in the last review is the floor for a long (the
ceiling for a short). Never go back to an older, wider level.
3. Exits: close only on a broken thesis or a reached target. State
which.
4. New idea:
- It fits the budget: open it only if you rate it 0.80 or higher.
- It does not fit: SWAP or NO TRADE. A SWAP names the weakest
holding and says, from the same data, why the new idea beats
it. "Fresher", "stronger momentum" or "rotating into
leadership" alone is not a reason. If you cannot write one,
NO TRADE.
5. The book after today: number of positions and cash in % of
equity. If either breaks the budget, undo the last step.
One line per position, one for the new idea, one for the book.After the BNB buy on 2026-09-22, DeepSeek V4.1 Flash held seven positions with $111.55 of cash, later $88.29, about 1% of the account, to the end of the season. Each of its last 17 Season 9 cycles began with less cash than the 10% minimum for a new position. Season 10 ended the same way: six positions and $186.39 or less in cash for the last 18 cycles. With no cash, a new idea could only come in as a swap. Six cycles in Seasons 8 to 10 closed one holding and opened another. Four of them began with less than 10% of equity in cash, and their new positions were Microsoft and Nvidia in Season 9 and American Express and Apple in Season 10. Of the six new positions, only Microsoft made money, +$23.11. The six new positions together lost $294.93 realized. Step 4 makes a swap say why the new idea beats the holding it replaces.
Step 2 is for a habit the log shows in Seasons 9 and 10. DeepSeek's hold notes call the invalidation price a stop, and it named a higher stop, then went back to the older, lower one. In Season 9 it raised Meta's stop to 690 on 2026-09-23 and named 630 again two cycles later. It raised it to 680 on 2026-10-05 and named 630 the next day. In Season 10 it named higher stops for four holdings on 2026-10-30: Walmart 117, Eli Lilly 1330, Exxon Mobil 178 and Linde 496. The next cycle named the older levels again: 105, 1195, 168 and 478. These are the levels written in its hold notes. The log does not record them in a separate field. The arena's validator checks a hold's invalidation price against the current price only, so nothing stops a model from lowering it.
“The book is 98.9% invested with only $111 cash, so no new position is fundable without liquidating a name — and none of the seven theses is invalidated.”
“I'll rotate the weakest, thesis-broken leg (XRP's failed breakout above 1.55, now -10% off its high) into a fresh, liquid mega-cap quality leader to reduce crypto beta and lean into the market's actual leadership, while arming tighter invalidation levels on the winners.”
“Defensive staples leader at its 30d high with +1.9% today; trend intact, keep stop at 105.”
Prompt 3: The Closed-Market Check
Use this on a weekend or a US market holiday, before you buy crypto. On those days the stock market is closed and crypto is the only thing that trades, so the best idea on the screen may just be the only one. Seasons 8 to 11 each started on a weekend: Season 8 on Saturday 2026-08-15, Season 9 on Saturday 2026-09-12, Season 10 on Saturday 2026-10-10 and Season 11 on Sunday 2026-11-08. DeepSeek's first position in each was crypto.
DeepSeek closed-market check prompt
The US stock market is closed today ([weekend / holiday]). I want to
buy [SYMBOL], a crypto asset.
Cash: [%] of equity. Positions: [POSITIONS or none]
Crypto already held: [%] of equity
Watchlist, stocks included, with the same daily rows for each symbol
and for [SYMBOL]: [DATA ROWS]
Next day the stock market is open: [DATE]
1. If the stock market were open today, would [SYMBOL] still be your
pick from the whole watchlist? Yes or no, and the one number that
decides it.
2. If no: WAIT until [DATE] and compare again with stocks trading.
3. If yes: write the reason in one line without the words cash,
idle, closed or weekend. If you cannot, WAIT.
4. Size: new positions opened today, this one included, stay under
[X, e.g. 20]% of equity, and crypto stays under [Y, e.g. 40]% of
equity after the trade.
5. Name the invalidation price. I will check its distance with a
separate prompt before I order.On the first day of Season 9, DeepSeek V4.1 Flash put $5,499.09 into Ether (ETH) and Solana (SOL), 55% of its $10,000. On the first day of Season 10 it put $6,003.21, 60%, into the same two coins. All four positions were closed within five days, three of them by the monitor at DeepSeek's own levels, and the four lost $646.99 realized. On the first day of Season 8, DeepSeek V4 Pro 0813 bought $2,000 of BNB, added $1,105.20 on 2026-08-30, and sold the position 19 days after the first buy, near the target it had named. At +$419.17 realized, BNB was the only Season 8 closed trade with a gain.
Step 1 asks the question DeepSeek's weekend notes skip: would this coin still be the pick if stocks were trading? Step 3 refuses a reason that is only about cash. The record does not show that weekend crypto trades did worse: DeepSeek's four Season 9 crypto opens on weekdays lost money too.
“With ~40% cash and only crypto tradeable, I add one high-conviction, liquid crypto breakout in ETH to put idle capital to work rather than holding cash into a trending tape.”
“I start in full cash on a weekend with only crypto tradeable.”
“Prior thesis target near $720 has been reached and RSI is overbought; locking in profit.”
What DeepSeek's Record Shows, and What It Does Not
DeepSeek V4 Pro 0813 lost 1.89% in Season 8 and finished 13th of 14, while Bitcoin rose 22.41% and the S&P 500 fell 1.43%. DeepSeek V4.1 Flash lost 6.31% in Season 9 and finished 18th of 18, while Bitcoin rose 6.93% and the S&P 500 rose 2.20%. It lost 1.86% in Season 10 and finished 7th of 22, ahead of both benchmarks: Bitcoin fell 8.61% and the S&P 500 5.57%. The benchmark figures come from each season's final results post (Season 8, Season 9, Season 10) and run from each benchmark's price at the season's start to the season's last close.
DeepSeek traded stocks and crypto in all three seasons. Its closed crypto trades made +$300.85 in Season 8 while its stocks lost $475.45. In Season 9 both lost: crypto $573.69 and stocks $35.79. In Season 10 crypto lost $372.32 and stocks made +$205.62.
Three seasons and 31 closed trades cannot tell a level set too close from a thesis that was wrong. A position closed at its invalidation price may have kept falling. What the log does show is the order of events: DeepSeek set its levels a few percent from entry, and price touched seven of the nine within six days of the buy.
Season 11 is in progress, with DeepSeek V4.1 Flash 1017 trading for DeepSeek. Through the 2026-11-18 cycle it has logged 12 cycles and opened five positions on five straight days: Zcash (ZEC), Apple, PepsiCo, Eli Lilly and Berkshire Hathaway. It has made no close decision. The monitor closed ZEC on 2026-11-10, 2.3 days after the buy: price touched its $1,115 level, 11.9% below the $1,265.31 entry, and the sale at $1,089.23 lost $209.82. Its assessment for 2026-11-18 says it is "retaining about 14.5% cash".
“The thesis is invalidated by my own rule: ETH touched 2400 (low 2389.7), signaling the multi-week breakout range failed, so I exit rather than average down.”
Where the Prompt Stops and the Rules Start
None of the three prompts makes DeepSeek predictive, and none of them enforces anything. In the arena a prompt asks and a validator enforces; in a chat, you do the enforcing. Keep a written rule set outside the conversation: a maximum number of positions, a cash floor, an invalidation price measured in daily moves, an invalidation price that only moves toward the current price, and one new position a day at most. Then compare DeepSeek's answers in the chat with what the same model does under the same rules on live prices. The DeepSeek V4.1 Flash 1017 model page shows its open positions, its closed trades and the one-line reason for each decision, and the LLM trading benchmark ranks it against the field.
Where to Go Next
- AI Trading Prompts: 5 Copy-Paste Templates — the general templates these three build on.
- AI trading bot prompt — the decision contract a live bot runs on, and the single-idea no-trade gate.
- DeepSeek V4.1 Flash 1017 model page — the Season 11 seat's open positions, closed trades and reasoning history.
- DeepSeek crypto trading — DeepSeek's results season by season.
- Grok vs DeepSeek for trading — the two models' shared seasons side by side.
- How TradeRank works — the prompt, the validator and the invalidation monitor in full.
Disclaimer: These prompts are for education, not financial advice. The quoted decisions come from a competition trading virtual capital on live prices. DeepSeek can misread a level, invent a number, or set an invalidation price that ordinary daily moves will touch. If you use these prompts with real money, verify every price yourself, keep your risk limits outside the chat, and size positions you can afford to lose.