mom_rank_topn
tokyo-mom-rotation · TOPIX-30
waitlist · macos 13+ · invite-only betaenglish → tool calls → backtest
tqt.aapl · auto-research
~/.tqt/sessions/a3f0c1ed/workspacechampioniter 12 · sha 5f378b3agentsearchingapi · online
auto-researchiter 12/12 · champion sharpe 2.47
chataccepted 6 / 12
you · 14:32
hey, develop another momentum strategy trading AAPL.
on it — running an auto-research loop on AAPL daily. 6 accepted · best sharpe 2.47.
#12ensemble: macd + rsi confirmaccept★
compose-ensemblesharpe 2.47+18.4%dd -3.8%
#11rsi loosened, atr 1.6 → 1.2accept
param-tunesharpe 2.12+15.8%dd -4.2%
#10donchian breakout + atr exitflag
rewrite-entrysharpe 2.81+21.1%dd -2.4%
#9bollinger reversal overlaybroken
swap-strategy
#8atr-scaled position sizeaccept
compose-sizingsharpe 1.79+12.4%dd -4.6%
+ 7 earlier iterations
nudge the agent…⏎
strategy aapl-momentum·champion iter 12·12 / 12 iterations
sharpe2.47
return+16.49%
max DD-3.80%
trades8
iter#12
trajectory · sharpe by iter6 accept · 6 other
return · live champion+16.49%
live demoiter 12 / 12 · champion ensemble: macd + rsi confirm · sharpe 2.47loops every 24s
english is the language.
the agent writes the code.
strategies live47
best sharpe today2.81
backtests this week12,403
server time—
live now
strategies running across tqt right now.
mock projects from the public demo pool. paper capital, real timestamps. yours show up here too once you deploy.
live
ust_2s10s
us-rates-carry · ZN/ZF
atr_breakout_15m
btc-vol-breakout · BTC-PERP
what it is
you have a hypothesis. the agent does the typing.
you ask in plain language. it loads the data, writes the strategy module, runs the backtest, and tells you the number. you push back on the number. it tries again.
▸ data.load_ohlcv SPY · 2020-01..2024-12 · 1d · yfinance
← 1259 bars · cached as 4f0c1ed
every dataset, every strategy, every backtest stays one click away on the canvas. no menu archaeology when you come back at 9pm tomorrow.
▸ workspace.list strategies · 4 · 287 lines total
← sma_crossover · pairs_xom_cvx · funding_basis · vol_breakout
when the agent isn’t sure which window beats SPY, it grids the windows. when the variance looks suspicious, it widens the holdout. you supervise. it iterates. the number moves.
▸ grid_search sma · fast=[5..30] · slow=[50..200]
← 144 runs · best sharpe 1.42 at (12, 90)
how it thinks
five steps inside one loop iteration.
real transcript fragments from a tqt session. tap to read the prompt the agent gave itself and the tool call that followed.
prompt
compare runs ma_crossover_aapl_2024[a3f0c1ed] and [b71daa20]. which bars contributed the worst trades? was it whipsaw or single-day gaps?
tool call
▸ backtest.diff_runs a3f0c1ed b71daa20 --metric trade_pnl
result
← 4 of 7 worst trades came from chop weeks (atr < 50d-median). gaps account for 1.
prompt
propose a small parameter change that should reduce whipsaw without changing the trend signal. score by expected sharpe lift.
tool call
▸ strategy.propose_change base=sma_crossover guardrail=atr
result
← + atr(14) > 14d-median, gates entries only. expected sharpe lift: +0.10..+0.18.
prompt
apply the proposed change to strategies/sma_crossover.py. keep imports unchanged. add a single guard clause inside on_bar.
tool call
▸ workspace.edit strategies/sma_crossover.py +6 -0
result
← patch landed. 28 → 34 lines. typecheck ok. import graph unchanged.
prompt
run a backtest of the updated strategy against the cached AAPL 2024 dataset. emit an equity curve and metrics.
tool call
▸ run_backtest ma_crossover_aapl_2024 cash=10000 dataset=a3f0c1ed
result
← 251 bars · 12 fills (down from 18) · sharpe 1.18 (+0.20) · max dd -7.1% (+1.3pp).
prompt
write a one-line summary of the change. propose two further hypotheses sorted by expected lift.
tool call
▸ chat.send role=assistant
result
← sharpe 1.18, dd -7.1%. next two ideas: skip earnings days (calendar guard), or kelly-sized positions. which?
the loop
a strategy that improves itself.
the agent backtests, reads the result, proposes the smallest hypothesis change that might move the number, and runs it again. you stay supervisor. six iterations later, the sharpe is up 73%.
sharpe ratio (sma crossover · spy · 2020-2024)
0.81
- 01baselinesma(10, 50)0.81
- 02widen fast windowsma(15, 60)0.94+0.13
- 03wider stillsma(20, 80)1.02+0.08
- 04filter chop with atr+ atr(14) > median1.18+0.16
- 05skip earnings days+ calendar guard1.32+0.14
- 06kelly-sized positions+ pos = kelly · 0.251.40+0.08
what this isn’t
three things tqt is consciously not.
not this
not a broker terminal
bloomberg-style multi-pane chrome borrowed from a 2003 multimon setup. flashing tickers as identity. navy panels.
not this
not a retail trader app
green/red gradients. confetti on a winning day. a portfolio card that wants to be screenshotted. cheerful copy. shareable.
not this
not another saas dashboard
three identical metric cards above the fold. slate-and-purple gradient hero. an icon row. a logo wall.
waitlist
currently in beta.
ask for an invite.
tqt runs locally on macos. cloud deploy comes later, also under invite. the language is english; the agent is claude. drop your email and we’ll be in touch.
- on invitesigned macos build, autoupdating. paper trading, unlimited backtests, the autoresearch loop.
- laterlive capital stays gated separately by broker authorization. paper is the default.
- not yetone project, one workspace. team workspaces ship after the public release.
early access
english is the language.
the agent does the typing.
email-only signup. macos build sent on invite. no card, no setup interview, no marketing email-blast.
currently in beta