Describe the idea in plain English. The agent writes the strategy, runs the backtest, reads the result and comes back with a sharper version. It has a real Python interpreter for the research behind it, and every step is logged in a session you can audit.
From a hunch to code to market evidence.
Describe your idea. The agent turns it into research, code, visualization, and tests across granular market data.
orb_15mESH6 15m orb_15m.py
test an opening range breakout on ES, and backtest on last month MBO data
Drafting orb_15m: long the break of the first 15 minutes, stop at range mid, flat by 14:30. Fills priced off L3 queue position, not mid.
- +
class ORB15(Strategy): - +
bars: Bars(interval_ns=60_000_000_000) - +
if close > orh: orders.ioc(side=BUY, qty=1) - +
stop = or_mid; flat_at = "14:30"
Details Previewing
- ├─Dataget_data ESH6 GLBX ohlcv-1s 2026-01-12ok
- ├─Backtestrun_backtest orb_15m v1ok
- ├─Fillsl3_queue ESH6 mbook
- └─Validatewalk-forward5/6
3.47M events, 21 sessions
Survives out of sample: +14.2%/yr net of fees and slippage, −1.8% max drawdown. Opening the run.
Ask Atlantive...
Claude Fable 5.1
6959.50
Equity, net of fees Jan 12 to Feb 10
Resultorb_15m v1
- Return / yr
- +14.2%
- Buy and hold
- +0.6%
- Max drawdown
- −1.8%
- Sharpe IS / OOS
- 1.71 / 1.34
- Profit factor
- 1.42
- Win rate
- 42.1%
- Walk-forward
- 5 / 6
the −1.8% drawdown, where did it come from?
Jan 29. CPI at 08:30: the gap opened through the range and the first pull never held. Three straight losers, one session. The guard skipped the print, not the session.
- ~
skip = ["cpi"] # not the session - +
skip_session("cpi") - +
calm = atr(5) < 2.0 * atr(20)
Details Applied
- ├─Recompiledsl to engineok
- ├─ReplayJan 12 to Feb 11ok
- ├─Diffv1 to v2−3
- └─Validatewalk-forward6/6
3.47M events, 21 sessions, unchanged
Session, 12 calls23.0 s wall
Diffv1 to v2
- Max drawdown
- −1.1% v1 gave back 1.8% on one session
- Return / yr
- +14.2 to +18.7%
- Trades
- 19 to 16
- Avg R
- +0.26 to +0.41
- Win rate
- 42.1 to 50.0%
- Sharpe OOS
- 1.34 to 1.66
- Compile
- 0.31 s
- Replay
- 8.90 s
why did my fill wait eight seconds?
Your order joined the queue at 6103.75 behind 31 contracts, fourteenth of twenty-seven. The tape reached you at 09:47:12.480. No mid price, no assumed fill: the book that was really there.
- ├─BookESH6 mbo, 09:47:04.118ok
- ├─Queuel3_queue, 14 of 27ok
- └─Fill8.362 s, 0.41 ticksok
Book, L309:47:04.118
6105.00312
6104.75266
6104.50198
6104.25147
6104.0041
Spread 0.25
6103.50ORH118
6103.25164
6103.00221
6102.75257
6102.50289
6102.25334
6102.00356
TapeTime and sales
09:47:04.118BUY36104.00
09:47:05.402SELL86103.75
09:47:06.930SELL26103.75
09:47:07.114BUY46104.00
09:47:08.667SELL146103.75
09:47:09.240BUY56104.00
09:47:10.881BUY26104.00
09:47:11.503SELL76103.75
09:47:12.480SELL16103.75
- Order
- BUY 1, LMT 6103.75
- Queue
- 14 of 27, 31 ahead
- Filled
- 09:47:12.480
- Waited
- 8.362 s
- Slippage
- 0.41 ticks, 4 bp
plot PnL by stop distance and position size
Sixty-three cells, one surface. The plateau at 15 minutes and 1.75R holds across its neighbours; the peak at 25 minutes stands alone. Drawn from the same tick source as every other view, so it cannot repaint.
- ├─Sweep9 stops × 7 sizes63
- └─Plotsurface, pnl(stop, size)ok
Surface, pnl by stop and size63 cells
pnl stop (R) size
Candles
Order flow
Equity, net of fees
show me what changed
Two lines. The CPI guard now skips the session instead of the print, and an entry waits for a calm tape: the five-bar range under twice the twenty-bar range. Recompiled in 0.31 s, no warnings.
- ├─Editorb_15m.py, +2 −1ok
- ├─Checkclosed vocabularyok
- └─Compile41 nodes, 0.31 sok
- ▾ strategies/
- orb_15m.py●
- open_drive_fade.py
- ▾ data/
- ESH6 GLBX ohlcv-1s
- calendar.csv
- ▾ tests/
- walkforward.py
- slippage.py
- ▾ runs/
- 2026-02-11_14-08✓
- 2026-02-09_09-31
- 1
from atlantive import BUY, SELL, Bars, Rolling, Strategy - 2
- 3
class ORB15(Strategy): - 4
# opening range breakout on ES, one contract, flat by 14:30 - 5
class Params: - 6
range_min: int = 15 - 7
target_r: float = 1.75 - 8
rvol_min: float = 1.4 - 9
- 10
class State: - 11
bars: Bars(interval_ns=60_000_000_000) - 12
atr5: Rolling(window=5) - 13
atr20: Rolling(window=20) - 14
- 15
def on_event(self, view, ev, orders): - 16
if not self.bars.update(ev): - 17
return - 18
close = float(self.bars.prev_close) - 19
self.track_range(close, self.p.range_min) - +
if self.session.skip("cpi"): # v2, was skip cpi ±30m - 21
return - ~
calm = self.atr5.mean() < 2.0 * self.atr20.mean() - 23
rvol = self.bars.volume / self.adv20 - 24
if view.position == 0 and close > self.orh and rvol > self.p.rvol_min and calm: - 25
orders.ioc(side=BUY, qty=1) - 26
self.stop = self.or_mid - 27
self.target = close + self.p.target_r * (close - self.or_mid) - 28
elif view.position > 0: - 29
if close <= self.stop or close >= self.target: - 30
orders.ioc(side=SELL, qty=1) - 31
elif self.clock >= "14:30": - 32
orders.ioc(side=SELL, qty=1) # flat, never overnight - 33
- 34
# the fill model is the book: l3_queue, queue position, not mid
- Target
- native engine
- Compiled
- 0.31 s
- Nodes
- 41
- Warnings
- 0
- or15
- first 15m, RTH
- rvol
- vol / adv(20)
- l3_queue
- fill model
- Guard
- session(cpi)
is it ready for the prop firm?
Six of six walk-forward folds. Five hundred Monte Carlo resamples, 5th percentile +9.4%. Slippage tripled, still +11.2%. Daily loss and drawdown caps held on every session. Deploy ready.
- ├─Rulesdaily 2%, dd 4%, targetheld
- ├─Walk fwd6 folds6/6
- ├─ResampleMonte Carlo, 500ok
- └─Stressslippage ×3ok
Equity, v2 over v1Jan 12 to Feb 11
Optimiser, range and targetdd < 4%, daily < 2%
Peak 25m, 2.5R, +24.1%, one cell
Chosen 15m, 1.75R, +18.7%, nine neighbours within 2 pts
Resultorb_15m v2
- Return / yr
- +18.7% v1 returned +14.2%
- Max drawdown
- −1.1%
- Sharpe IS/OOS
- 1.94 / 1.66
- Profit factor
- 1.64
- Win rate
- 50.0%
- Trades
- 16
- Daily loss cap
- < 2%, held
- Max drawdown
- < 4%, held
- Profit target
- met
- Monte Carlo
- 500 resamples
- 5th percentile
- +9.4%
- Slippage ×3
- +11.2%
- Walk-forward
6 / 6
Objectives
Robustness
Most backtesters fill your order at a price the market never offered you. Here your simulated order stands in the recorded queue, behind every order that was really ahead of it, and fills only when the tape reaches it. No look ahead, no invented prices, replayed at 240× speed.
Candles with your trades printed on them, order flow, equity, all drawn from tick data that never redraws itself. When the view you need does not exist, describe it and the agent builds it, up to a full 3D surface of profit across every stop and every size.
A strategy is a short Python file, about thirty lines you can open, edit, diff and version. Change one line and it recompiles in a third of a second. The same file compiles to Pine Script when you want it on TradingView, verified to trade identically.
Before a strategy touches money it faces the rules prop firms apply to the traders they fund: daily loss cap, max drawdown, profit target. Then walk forward tests, five hundred Monte Carlo runs and slippage stress. A 63 cell sweep keeps the plateau that survives everywhere, never the peak that got lucky once.
Agent
Describe the idea in plain English. The agent writes the strategy, runs the backtest, reads the result and comes back with a sharper version. It has a real Python interpreter for the research behind it, and every step is logged in a session you can audit.
Backtesting
Most backtesters fill your order at a price the market never offered you. Here your simulated order stands in the recorded queue, behind every order that was really ahead of it, and fills only when the tape reaches it. No look ahead, no invented prices, replayed at 240× speed.
Charts
Candles with your trades printed on them, order flow, equity, all drawn from tick data that never redraws itself. When the view you need does not exist, describe it and the agent builds it, up to a full 3D surface of profit across every stop and every size.
Strategies
A strategy is a short Python file, about thirty lines you can open, edit, diff and version. Change one line and it recompiles in a third of a second. The same file compiles to Pine Script when you want it on TradingView, verified to trade identically.
Validation
Before a strategy touches money it faces the rules prop firms apply to the traders they fund: daily loss cap, max drawdown, profit target. Then walk forward tests, five hundred Monte Carlo runs and slippage stress. A 63 cell sweep keeps the plateau that survives everywhere, never the peak that got lucky once.
Write it in Python.
Export it as Pine Script.
A strategy is a compact Python class written in our trading DSL. A closed vocabulary, checked before execution, with invalid syntax rejected and corrected upfront.
When you want to use it on TradingView, the same file compiles directly to Pine Script. No AI guessing, no manual rewriting.
strategy.py
MeanRevert.pine
Years of futures data, built in.
Bars from 1 second to 1 month since 2010, plus one year of trades and quotes and one month of full order book data. Nothing to buy, connect, or import.
OHLCV bars
6104.00 6101.50
16 years
1 s to 1 mo
Trades
TimePriceSize
1 year
every print
Top-of-book quotes (L1)
Ask 42 Bid 62
1 year
every quote
Order-book depth (L2)
1 month
every level
Market by order (L3)
You, 14th of 27
1 month
every order
Illustrative previews of historical data.