AI assistant
Ask about the chart, build indicators and strategies, and research them with the quant lab.
The assistant lives in the panel on the left (Alt+A). It sees the chart you are on, reads the market, writes Flowscope Script, puts indicators and strategies on the chart and tests them. Connect it once with a Claude or ChatGPT plan, an API key or any OpenAI-compatible server; the connection stays on your device.
It never places, changes or cancels orders and never moves funds.
What it can do
| Area | Tools it uses |
|---|---|
| Understand the chart | the chart’s market, interval, indicators, scripts, drawings and position; a picture of the chart |
| Read the market | bars of any market and interval, the order book, large trades, liquidations, funding, open interest |
| Write scripts | the scripting docs and built-ins, a compiler check, putting the script on the chart (typed into the script editor as you watch, saved under My scripts) |
| Test strategies | the Strategy Tester report, over the last days or a date range |
| Change the chart | add or remove built-in indicators, switch market, interval or chart type, draw levels and notes |
| Research (quant lab) | the tools below |
The quant lab
Research that would take the assistant dozens of steps and long lists of bars runs inside Flowscope in one step, in parallel on all cores, and comes back as a short summary. Answers are faster, and a conversation uses far fewer tokens.
| Tool | What it answers |
|---|---|
Market profile (market_stats) | How a market moves: volatility, the size of a typical bar, fat tails, whether moves continue or revert (autocorrelation, variance ratios, Hurst), how efficient trends are, which UTC hours and weekdays move and trade most, the biggest bars. |
Signal study (signal_study) | Whether an idea has an edge before a strategy is built: what the price did 1, 3, 5, 10 and 20 bars after a signal, against every bar, with hit rate, t-statistic and the best and worst excursion on the way. |
Optimize (optimize) | The best inputs of a strategy: every combination (or a fixed random sample when there are too many), each a full backtest. The search only sees part of the bars; its best runs are then tested on bars it never saw (holdout), or the choice is repeated fold by fold on the past only (walk-forward). It reports how much each input matters and whether the best inputs sit on a plateau (their neighbours do almost as well) or a spike (likely overfit). |
Monte Carlo (monte_carlo) | How much of a backtest was the luck of the order: the strategy’s trades resampled thousands of times, with the range of returns and drawdowns, the chance of a loss and of a deep drawdown, and the longest losing streak to expect. |
Compare markets (compare_markets) | Whether the edge is general: the same strategy and inputs on several markets and intervals at once. |
Each study runs on the bars of the chart’s market and interval unless you name others, over the strategy’s test period or the days you ask for; it loads older history first if the period needs it. A strategy runs exactly as on the chart, with its inputs and its properties (capital, sizing, commission, slippage, margin). Ask in plain words, for example:
- “Profile ETH on 15 minutes over the last 30 days.”
- “Is there an edge in buying when RSI drops below 25 on BTC 1h?”
- “Optimize the fast and slow length of my strategy with walk-forward, then run a Monte Carlo on the result.”
- “Does this strategy also work on SOL and ETH, on 1h and 4h?”
With “apply the best inputs” the assistant writes the chosen values into the strategy’s inputs on the chart; you see the Strategy Tester update.
Reading the results
- Out of sample counts. A search will always find something that looked good in the past. Trust inputs that also hold on the held-out bars, on neighbouring values and on related markets.
- Few trades say little. Runs with fewer trades than the minimum (20 by default) are not ranked.
- Margin calls. Sizing at or near 100 % of equity with costs makes the backtester close small parts of the position for want of margin, as TradingView does. Each of those counts as a trade, so the trade count rises and the win rate sinks. The reports say when this happened; size at most about 95 % of equity for clean figures.
- Results are measurements of the past, not advice or a forecast.