How Tradytics Signals are built: the research behind them
By Faizan Ahmad. Updated .
Tradytics Signals come from one of the largest studies of retail options flow published by any platform: more than half a billion pattern and horizon combinations tested on 10.9 million options prints over 378 sessions, joined to real 1-minute stock bars and real option quotes, with the filter judged on months it was not fitted on. Most of what was tested failed. Only patterns that held up became signal patterns, and a trained filter hides the weakest prints. A signal is a real order that matches one of them, and the trade is to buy the same contract, at /signals.
What data went into the study?
| Input | What it was |
|---|---|
| Options prints | 10.9 million single-leg prints, with spreads, multi-leg orders and late reports left out |
| Sessions | 378, from May 2025 to October 2026 |
| Stock prices | Real 1-minute and daily bars for every name, with dividends added back into returns |
| Option prices | Real option quotes (the national best bid and offer), and settlement at the official close on expiry day |
| Combinations tested | More than half a billion pattern and horizon combinations |
Spreads were left out because a spread's legs can't be read one at a time: a call bought as half of a spread isn't a bullish bet on its own. How to read spreads in options flow explains why.
How were the patterns tested?
Two ways, because they answer different questions.
- Did the stock move? Each pattern was searched on one period, judged on a second, and scored once on a third that the search never saw. Patterns had to point the right way in all three, and a price-only control showed how much came from the flow itself rather than the stock's own setup.
- Did the contract make money? This is what an options trader actually holds. Each pattern bought the printed contract at the ask, paid fees, and was followed on real prices. One signal per ticker per session, so a burst of repeat prints in one name can't count many times.
The filter that hides weak prints was trained only on earlier months and asked to judge the next, so every one of its calls was out of sample.
What failed?
Most ideas about options flow didn't survive real prices. Publishing what failed is as useful as what worked:
- Stock direction over 2 to 10 days. Patterns that looked strong did no better than a control with random signs.
- Big intraday edges. The striking ones came from the same ticker printing again and again on one day. Counted once per ticker per session, they shrank to almost nothing.
- Stock moves too small to trade with options. The stock-direction patterns that held up moved the stock a fraction of a percent, which doesn't pay for an option's spread and time decay on real quotes.
- Chasing what worked lately. Taking the patterns that worked over the last 60 or 90 sessions and trading them next did no better than random picks.
- Following a stock's own flow direction. A ticker's lean carried into its next window about as often as a coin flip.
- "Unusualness" scores. A ticker's top 1% of premium, the top 1% of size against open interest and similar rankings did no better than the base rate.
- Machine learning as a picker. Models trained on contract outcomes couldn't pick winners on average, though they were good at spotting losers. That made them a filter, not a source.
- Patterns that faded in a later replay. Late-day calls in very cheap stocks, small-cap call buying and record-size prints all looked promising in the search and were dropped after a replay on later sessions.
Earlier Tradytics work came to the same place for gamma walls, pinning, gamma direction, net flow direction and whale direction: on their own they don't predict the next move. That's why Tradytics shows them as positioning, not as Signals, and why following unusual options activity on its own is a weak plan.
What survived?
Seven patterns held up on the contract itself and became signal patterns:
| Pattern | What it watches for |
|---|---|
| Repeat flow | The same at- or out-of-the-money contract bought again and again in one session, with little sold against it |
| Opening calls | New call positions about as large as the contract's open interest, bought while implied volatility sits at the stock's normal |
| Heavy calls | A call's volume running far above its open interest in a low-priced stock |
| ETF repeat | The same in-the-money ETF call bought again across recent sessions |
| ITM puts | Puts bought at least 2% in the money in liquid, high-priced stocks |
| Call burst | A sudden burst of bullish premium in a stock on a strong market day |
| Index tide | Calls bought in single stocks while SPY, QQQ and IWM flow leans heavily bullish |
Each one, with its exact conditions, is in the seven patterns behind Tradytics Signals.
What filters run before a signal is shown?
- The trained filter. A model scores every print that matches a pattern and hides the weakest 40%. What it learned was mostly about the contract (moneyness, calls or puts, premium size, liquidity), not about how exciting the flow looked. Repeat flow has its own bar instead: the seventh buy on the same contract in a session.
- Price must confirm the side. A buy has to print at or above the ask.
- Rich options are dropped. Implied volatility above 1.3 times the stock's normal, at-the-money contracts under $0.25, and 0DTE buys in the last 30 minutes of the session lost money in every group and never become signals.
- Tight markets on three patterns. ITM puts, heavy calls and opening calls also need a bid/ask spread of at most 10% of the midpoint.
Why is there no AI confidence score?
Because nobody can measure one. A language model asked "how confident are you?" returns a number that isn't tied to any record, and it reads like a probability when it isn't. Tradytics shows something you can check instead: the order that fired the signal, and how the contract has done since, marked every 5 minutes and settled at expiry.
Rules are frozen. Changing any condition makes it a new pattern, which needs its own replay on past sessions before it can produce signals.
Common questions
Are Tradytics Signals made by AI?
No language model picks them. The patterns are fixed rules that held up in the study, and a trained filter hides the weakest prints. A language model only writes each signal's short note from its facts.
How big was the Tradytics options flow study?
More than half a billion pattern and horizon combinations tested on 10.9 million options prints over 378 sessions, joined to real 1-minute bars and real option quotes, with the filter judged on months it was not fitted on.
Why do signals use the contract's result and not the stock's?
Because the contract is what you hold. A stock can move the right way by less than the option's spread and time decay, and the trade still loses. The study counted what the option did, at real prices, and each signal's Return does the same.
When can I see signals?
Plus and Pro see signals live. Free sees them a day late. Plans are in Tradytics pricing.
How are signals different from the AI Fund?
Signals come from options flow patterns and are single calls and puts. The Tradytics AI Fund is a fundamental long/short options book run by AI agents on dated catalysts, and doesn't use flow as evidence.
Sources
- Options Price Reporting Authority (OPRA): the consolidated last sale and quotation feed from the US options exchanges, the raw material for the prints and quotes studied.
- Selina Han, "Unveiling the Sophistication: Understanding Retail Investors' Trading Behavior in the U.S. Options Market" (Cboe, May 17, 2024): Cboe's study of how retail investors trade options.