Invite-only access
Choose the market, the features and the algorithm. Tune the hyperparameters yourself and set what the model is rewarded for. Backtest it, run it on paper for as long as it takes to believe it, and promote it to real money only when you say so — one approved order at a time. The quant curriculum that teaches all of it ships with the platform.
Invite-only while we scale. Have a code? Sign in
We email you once, when your invite is ready. By joining you agree to our Terms and Privacy Policy.
{
"name": "MA cross, RSI filter",
"entryRules": [
{
"indicator": { "type": "SMA", "params": { "period": 12 } },
"operator": "crosses_above",
"compareWith": { "type": "SMA", "params": { "period": 26 } }
},
{
"indicator": { "type": "RSI", "params": { "period": 14 } },
"operator": "less_than",
"value": 70
}
],
"exitRules": [
{
"indicator": { "type": "SMA", "params": { "period": 12 } },
"operator": "crosses_below",
"compareWith": { "type": "SMA", "params": { "period": 26 } }
}
],
"entryCombine": "AND",
"exitCombine": "AND"
}Two ways to trade
Plenty of profitable strategies are a moving-average cross and a risk limit — you should not need a neural network to run one. And plenty of edges cannot be written down as a rule at all. Do either here, or run both side by side and let the backtests settle it.
Classic systematic trading. Build the strategy yourself from indicators and conditions — no machine learning involved, and no training time.
Deterministic. You can trace exactly why it fired.
When the pattern is not something you can write down as a rule, train a model to find it on your data instead.
Adaptive. Learns behaviour you did not have to specify.
Both deploy the same way, inherit the same safety rails, and produce the same audit trail. The regime orchestrator can switch between them as the market changes — a rules strategy in chop, a trained model in a trend.
Full control
Most tools hand you a strategy and hide the machinery. This one hands you the machinery. Every item below is a parameter the trainer actually accepts.
Your symbols, your timeframes, your history depth. Datasets are versioned, so a result you liked stays reproducible.
Technical indicators, fundamentals, news sentiment, or a mix. If you think volume profile matters and momentum does not, build it that way.
Pick the algorithm, then tune it. Learning rate, batch size, steps, network shape — exposed, not buried behind a "smart defaults" toggle.
Reinforcement rewards are weighted across four components — returns, downside, differential and Treynor — and you set the weights. Default is an even 25% split; tilt it toward downside if capital preservation matters more than upside.
Maximum position concentration, reward scaling, and the drawdown at which a live deployment flattens and stays down.
Trained models are yours. Keep them private to your account, publish them to the model zoo, or delete them outright.
Three families of model
You are not picking a product tier here — you are picking a hypothesis about how your market works.
An agent that learns a trading policy by trading — thousands of times over your history, keeping what paid.
You shape the reward: how much it should care about raw return versus downside risk, and how large a position it may ever hold.
A sequence model that predicts where price goes next, from the features you decide are predictive.
You choose the architecture, the lookback window, the horizon, and which indicators become inputs.
A model that turns news and social text into a signal you can trade or use as a feature in another model.
You pick the sources and the window, and decide whether sentiment trades directly or feeds a price model.
Learn it while you build it
Not a blog and not a video course bolted on for marketing. It teaches the things that actually decide whether a model survives contact with a market — and it is written by the people who built the engine you are pointing it at.
10 modules
8 modules
9 modules
The platform
No notebooks to babysit, no VPS to keep alive, and no scripting dialect invented for one vendor.
Pull historical bars for a symbol and timeframe, stored as a versioned dataset with a market-regime label attached.
US equities and crypto · 1Min–1Day
Reinforcement learning, price forecasting, or sentiment. Jobs bill per second of actual compute, and your tier includes an allowance — you see the estimate before it runs.
Serverless, per-second billing
Measured against history before it can touch an account, and segmentable by regime — so a strategy that only works in trends says so.
Regime-segmented
Strategies are plain JSON. No black box, no scripting dialect invented for one vendor, and you can diff two versions.
Validated before deploy
Deploy and the engine evaluates on every closed bar, places the order, tracks the position. Stocks park when the market closes and resume on their own; crypto runs around the clock.
Every 60 seconds
Every real-money order waits for your approval — an SMS reply or a signed email link. Nothing fires while you sleep.
Hedge tier and above
Map strategies to trending, sideways and volatile markets; the orchestrator classifies the live regime and routes accordingly, logging every switch.
Quant tier and above
Every evaluation, order, route change and halt writes an event with its reason.
Exportable event feed
How it works
Most users live at step four for a long time. That is the point — paper trading costs nothing but time, and time is what tells you whether a strategy is real.
Choose a symbol and timeframe. Bars land as a versioned, regime-labelled dataset.
Pick RL, forecasting or sentiment. Watch GPU progress and cost as the job runs.
Measure it against history before it can touch an account. Keep it or discard it.
The engine evaluates on every closed bar and places paper orders automatically.
Promote it when you trust it. Each real-money order waits for your approval.
Breach your max-drawdown limit and the strategy flattens and stays down. It never restarts on its own.
Three consecutive failures and the deployment stops and tells you, instead of quietly retrying into a moving market.
Paper deployments are refused a live brokerage key outright, so a misconfiguration cannot spend real money.
Where we fit
An honest comparison — plenty of people are better served by one of the other three columns.
| CodonQuant | Signal services | DIY notebooks | Charting platforms | |
|---|---|---|---|---|
| Models you trained yourself | Yes | No | Yes | No |
| GPU training included | Yes | No | Bring your own | No |
| Backtest → paper → live in one place | Yes | No | Partial | Partial |
| Strategy frozen at deploy | Yes | — | No | No |
| Per-order approval on real money | Yes | No | No | No |
| Decision-level audit trail | Yes | No | Bring your own | No |
| Runs without your machine on | Yes | Yes | No | Partial |
We are built for one person in particular: you want your own models, trained on your own data, without also running your own GPU fleet, scheduler and execution plumbing.
Scope, stated plainly
Every platform makes these choices. Most leave you to discover them after you have paid, so here they are up front.
The workbench is designed for a real screen — charts, backtests and training runs side by side. Live-order approvals come to your phone by SMS, so the time-critical part travels with you.
Two markets, covered properly: equities through the session calendar, crypto around the clock. Forex and options are outside our scope.
Trading runs through your own Alpaca account. Your keys are encrypted at rest, and we never hold your funds or take discretion over them.
We onboard deliberately because we route real orders against real accounts. Joining the waitlist locks in founder pricing for when your invite lands.
Pricing
Every plan includes the full workbench. What changes is how much you can train and whether real money is on the table. GPU hours beyond your allowance bill as metered usage — you see the estimate before a job runs.
Kick the tyres.
$0
Included with your invite
For one serious strategy.
$29/mo
First tier with real money.
$99/mo
Running a book.
$299/mo
Running a desk and need more than Institutional? Tell us what you need — Enterprise is a conversation, not a checkout page.
Access is invite-only while we scale the training fleet — join the waitlist and you keep founder pricing for life once your invite arrives. There is a feedback button on every screen in the app, and it reaches an engineer, not a ticket queue.
We email you once, when your invite is ready. By joining you agree to our Terms and Privacy Policy.
CodonQuant is software for building and running your own trading strategies. It is not investment advice, and we do not manage money or recommend trades. Automated trading can lose money, including more than you intended — backtested results describe the past and promise nothing. You connect your own brokerage account and you approve every live order.