AQ Genesis
It opens from Strategies → 🧺 AQ Genesis. You give it a Symbol and a Direction (Long / Short / L+S), press START and it sets off researching on its own, without stopping: it generates and filters strategies while you sleep. The Winners' Room fills up in real time with the best ones, sorted by CONFIDENCE, not by profit — the promise is a basket of robust strategies, not the most profitable one in sample.
The wall: a ladder that cuts at the first failure
Every strategy climbs 7 rungs; the moment it fails one it is discarded and the following ones are not evaluated (saving compute). Each rung has a checkbox to turn it on or off on the fly — untick one and you will see instantly how what passes the filter changes. The ✓✗ ladder in the «Last decision» panel shows where each strategy dies.
⚙️ Settings: you set the bar
The ⚙️ Settings button in the header opens a window with everything that decides what Genesis finds. Every field carries its explanation beside it: you do not need to know the theory to change it sensibly. Changes are saved and applied to the next investigation — moving a bar mid-run would make the strategies sealed before and after it not comparable with each other.
Minimum number of trades, maximum PBO, AQ WF Score, AQ Survival Score, the quick filter's Profit Factor, the margin over buy-and-hold and on which markets robustness is tested.
How the score that ranks the Room is weighted between Walk-Forward, Survival and anti-overfitting. They count as proportions: they do not have to add up to anything in particular, and the window shows you the real split in %.
Maximum number of entry conditions, TP and SL ranges, candidates per round, the Room's cap, the shift [n], the exit criterion and the regime filter.
The three «best one wins» thresholds — the only ones applied instantly, even with the research running — plus the point value and the commission.
Regime filter: searching in one market environment only
In ⚙️ Settings, next to the Exit criterion, the Regime filter (AQL code) field takes a fixed condition that every strategy Genesis finds must meet. For example, to search for entries only with a low VIX:
It is added to every candidate's entry and travels inside every sealed strategy. It does not count as a searched condition and does not raise the chance bar. Condition names already state the series: CLOSE[0] of $VIX < 20.
Below the field you see whether the filter is valid and how many conditions it has; with errors, the investigation will not start. It is remembered for next time. Empty = no filter.
Rung 4's always in the market version becomes always whenever the filter holds: the strategy has to add something on top of the regime itself; a good regime is not enough.
The filter can read another series with of (symbol in quotes, registered in the Quote Manager) or another timeframe with @. The search then runs on the CPU: slower.
The syntax is the AQL Language's (see "Reference series"). You can join several conditions with and.
The context strip: never launch blind
Right under the header there is a line summarizing, always in sight, the settings the research will run with: $50/pt · $1.00/trade | Exit: TP/SL | Shift 0-10/5 | 12 markets. It matters more than it seems: the exit criterion completely changes what Genesis finds — the best entry depends on how you get out — and a run can last hours. A click on the strip opens Settings.
Rung 4: edge or market? — separating timing from the rise
In a rising index almost any long «works» by riding the trend. Rung 4, «No edge vs market», compares each candidate with its «always» twin (same exits and same order, entering at any moment) per day in the market, year by year and without the best year. It discards the red ones in training (cheap filter) and in validation (the wall, before Walk-Forward). It applies to longs and shorts. Full detail in Edge or market?
There are no margins to tune: it is switched on and off with reason 4. It replaces the old margins over buy and hold and over «always in the market», which compared RoMaD (profit over drawdown). RoMaD does not measure time in the market and rewarded the strategy that happened to be out during the big fall: of the 100 strategies sealed in a test on daily @ES, 45 passed those margins and only collected the market.
In the Winners’ Room, each sealed strategy shows three dots: training, validation and reserve (green, amber or red). The reserve only informs. The OOS always column still tells you, in money, what the twin would have made in the out-of-sample stretch; if the strategy made less, its OOS is shown in amber.
Comparative curve chart
Below the tables, a chart overlays up to 3 equity curves: Buy & Hold (gray, the reference), Best strategy (green) and the Selected one (orange) — click any row (in Winners or Discards) to see its curve; click again to remove it. A $ / % button switches between the real amount and a comparable shape, and checkboxes show/hide B&H and the Best. That way you SEE the edge, you do not just read it.
«Best one wins»: 100 different ideas, not 100 clones
A genetic algorithm converges and rediscovers the same strategies over and over. Genesis de-clones them on two levels: the EXACT clone (the same recipe) is discarded instantly; and the FAMILY near-clone — similar but distinct strategies, with a monthly return correlation ≥ 0.90. When a sealed strategy is a near-clone of a sister already present, it only replaces it if it improves substantially in Confidence (+4) OR in Profit (×1.15); otherwise it is redundant and gets discarded — because among near-clones confidence barely varies, so profit breaks the tie. That way the Room holds 100 different ideas (or however many you set in ⚙️ Settings), each in its best version.
- →Min. correlation (0.90): how similar they must be to count as a «family».
- →Confidence jump (4): how much more confidence is needed to replace.
- →Profit factor (×1.15): how much more $ is needed to replace.
Nothing is lost: log, viewer and analysis
While researching, Genesis writes to disk EVERY strategy it analyzes (the discarded ones and the clones too), in its own folder C:\AniQuant\Genesis. It is not just the Room: it is the complete logbook of the research, to be mined afterwards.
From the Room to your trading
Right-click a sealed strategy — in the live Room or in the Viewer — to open the Actions menu, the same one as in the other five windows: Performance, AQ Contrast, Survival Test, Walk-Forward, optimize in the Lab or just the exits, edit the AQL, save to the Library, export to the six platforms and add to the Portfolio. It is explained in full here.
In the viewer only the sealed ones are actionable (they keep their full genome in a side file); the mass discards are telemetry only.
Choosing by hand which strategies go to the Portfolio
When a run finished, the menu only offered two extremes: send this one or send the whole Room — and the Room usually holds a hundred. There was no way to keep just the six that interest you.
The Winners' Room and Best discards carry a checkbox at the start of every row; the header one ticks or unticks all the visible ones at once. Next to each grid, a counter of how many you have ticked, so you always know how many you are sending.
«Send N selected to the Portfolio». If none is ticked, it changes to «Send this one to the Portfolio» and sends the row you right-clicked, as always.
The ticks survive re-sorting by any column and the repainting while the research keeps running; they are cleared when a new one starts. The discards can also be sent — Genesis was already saving their genome, all that was missing was being able to ask for them —, and if something ticked cannot be sent, you are told how many and why, instead of failing silently.
How do I see whether the strategies Genesis found resemble each other?
The Winners' Room sorts them by confidence, but it does not compare them with one another. And that matters: ten excellent strategies that win and lose at the same time are not a portfolio, they are a single strategy repeated ten times. To see it, you have to take them to the Portfolio.
- In the Winners' Room table, right-click any row — you do not need to select anything first.
- Choose «Send the whole Room to the Portfolio».
- Open Strategy Portfolio and press Run.
- Go to the Correlation tab.
Red = they move together (redundant). Green = they diversify.
⚠️ The Portfolio's numbers will not match the Room's, and that is not a bug: Genesis shows the training stretch and the out-of-sample (OOS) stretch separately, whereas the Portfolio runs over the whole history. They are two different things.
Close it and carry on tomorrow: the persistent Room
The Winners' Room saves itself (one snapshot per Symbol + TF + Direction) every time it changes and when you stop. If you switch the PC off or close Genesis, you do not lose what was sealed. When you pick a market you will see a notice — 💾 Room saved: N sealed — and with the Resume Room box ticked, pressing START re-populates the Room and carries on researching from where it was. Each direction (Long / Short / L+S) has its own snapshot.
Every strategy knows where it came from: Origin
Every strategy now carries an Origin label — the tool that created it: Genesis, Lab, AQL or Seasonal — visible as a (sortable) column in the Library and the Portfolio. The ones that get re-packaged (Basket, Multi-TF) inherit the real origin, so you know a strategy inside a basket was born in Genesis, not a generic «Basket».
Explore EVERYTHING, not just moving averages: balance by group
There was a natural bias: because the Price group has many indicators (SMA, EMA, channels…) that combine with each other, the genetic algorithm almost always produced price strategies and the oscillators barely showed up. Genesis now shares out the exploration:
Each one explores ONE group, rotating (Price, Oscillators, Volatility, Volume…). The groups with more indicators (Patterns, Price) get more turns.
The full pool, to keep generating cross combinations (RSI < 30 AND SMA < channel).
The 🔬 Exploring now KPI in the mission panel tells you live which group is being researched at any moment.
And within each group, first the indicator and then its variant. Without this, the moving averages (EMA, SMA), with thousands of parameter combinations, came out 60 % of the time and a candlestick pattern hardly ever. Now the genetic algorithm first picks which indicator and then with which parameters: the least used one is tested about 7 times more than before.
First pass: every group, starting with those with the most untested indicators. Before any group repeats, each one gets its round, and the one with the most indicators not yet tested goes first (Patterns with 35, then Price, Time…). If a group does not fit whole on the graphics card, it is explored in chunks and each chunk starts with the missing indicators. With minute data (about 20 seconds per round) this full pass takes a few minutes; at the end of every run the log says how many indicators of the pool were tested. In a 10-minute run on 3 years of 1-minute @NQ, all 86 were tested.
You choose which indicators it researches: «Indicator usage»
In the left column, under «Discarded by reason», the Indicator usage — live card appears when the window opens, with every candidate indicator and a checkbox on each row.
Indicator · Group · Evaluated (strategies that contain it) · % eval. · Sealed · Yield = % of the Room ÷ % of what was evaluated. In green (≥ ×1.5), the genetic algorithm undervalues it; in red (≤ ×0.67), it is tested a lot and wins little. Click a column to sort; it refreshes every 1.5 s.
Unchecked indicators are not used in the next research run, and the choice is remembered when you close. Buttons Check all, Uncheck all and Only the least used (keeps checked the less tested half of the last run). If no usable indicator is left, it researches with all of them.
With Record log checked, when it stops it also saves Genesis\uso_indicadores_<symbol>_<date>.csv, and the log summarises how many indicators in the pool were tried and which were never tried.
Study first and validation 1.2.0
Checkbox next to «Resume Room», ticked by default. Before starting, AniQuant runs its labs (TP/SL map, seasonality, duration profile, edge scanner…) on the search period, without looking at the reserve, and adds what it finds to the Genesis pool; its exits (the map’s TP/SL and the exit after N bars) rotate by round with the Genesis one. It takes about one more minute. In the log: «Preliminary study: N findings…».
Settings → «Validation (% of what the search sees)» now defaults to 0: the genetic engine learns from the whole period and the wall judges with its own tests (Walk-Forward, Survival, PBO, DSR); the reserve remains the clean exam. In out-of-sample tests (ES and gold, 2021-2026), 25% hurt: it took the most recent years away from the genetic engine. It can be raised from 5 to 50.
Seed, tabu memory and PBO
Settings → The search. 0 = random (default): each research run explores different combinations, so repeating it another day does add (with «Resume Room» it accumulates). A number = fixed: the run repeats identically. The log states the seed used at start-up, so you can repeat it.
Settings → The search, off by default. When on, the genetic algorithm does not propose strategies already analysed again. It helps explore new ideas, but the Room fills more slowly and tends to lose the most profitable ones.
PBO is measured per batch: the round's candidates plus 60 strategies picked at random from everything tested. Measuring it with the best ones only always gave a high PBO even when they had a real edge, because they were chosen on the same data. A batch with fewer than 4 distinct strategies waits to join the next round. If it could not be measured, the PBO column shows «—» and Confidence is computed without it.
It sits on the same rung as PBO: a strategy is rejected if the batch PBO is high or if its own DSR is low. DSR looks at its Sharpe and discounts the luck of having tried thousands of combinations. Tested on a real market against its shuffled copy, 0.20 filters out most of what looks good by chance. The textbook 0.95 lets nothing through, because the genetic search tries very similar strategies. Setting 0 disables it.
It was the threshold that best separated the real market from the shuffled one: with 65 (it was 55) more sealed strategies win out of sample and less noise gets in. Raising the minimum Profit Factor, on the other hand, made things worse: a very high PF in the search is usually overfitting.
If a sealed strategy is in the market 75 % of the time or more, the Hall flags it in amber next to its name (for example, ⏱ 97 %). In a rising market such a strategy makes money like buy-and-hold with a stop: neither the search period nor the OOS can separate its edge from the market rise. It is not rejected; check it with AQ Market Shuffle.
Checkbox next to «Log to disk». When you stop (or at 2/3 of the time limit), Genesis repeats the same search —same settings and same seed— on a shuffled copy of the market for half the time: same volatility and same rise, no pattern. What it seals there does not enter the Hall; it is only counted. The «Random control» card compares, per direction, the sealed strategies per 100,000 evaluated: if the real market seals 3 times or more than the random one, there is an edge; between 1.5 and 3, doubtful; less, it cannot be told apart from chance, and the sealed strategies in that direction carry the 🎲 mark. It is the only test that separates a strategy that merely rides the market rise from one with a real edge.
On daily data HourOfDay and MinuteOfDay are not proposed (on daily bars they are a constant), and near-clones are summarised per round in the log instead of being written one by one. If the window is narrow (a laptop with Windows scaling), the header wraps onto two rows and the START button always stays in view.
Conditions with sense
In ⚙️ Settings, the Conditions with sense row has three settings that help Genesis search where it makes sense. They only apply to Genesis: the Lab tries exactly what you ask it to.
By default, the range of each threshold (the 35 in «RSI < 35») comes from the percentiles of the indicator in the search data, from 10 to 90, not from a fixed table. So a Momentum is tried in ES points or in EURUSD pips, each on its own scale, and no thresholds that are never met appear. It is computed with several periods at once, because a 2-bar RSI moves much more than a 30-bar one, and it never looks at the final reserve. With extremes (5 to 95) rare values also come in. From the catalog goes back to the usual ranges. The log says at the start how many indicators were calibrated and gives an example.
Each round adds up to 24 conditions made with that round's indicators, which the genetic algorithm combines with the rest:
- For N bars: Count(RSI(Close, 14) < 35, 3) >= 3
- Changes direction: (MFI(21) > MFI(21)[1] and MFI(21)[1] < MFI(21)[2])
- Against its own average: MFI(14) > SMA(MFI(14), 20)
They are ordinary AQL formulas: they run on the GPU, are saved, can be edited and are exported to all six platforms. They are not known strategies, just ways of looking at an indicator.
Many trading ideas work in two steps: first something arms the entry and then something else triggers it. With the box ticked, each round adds up to 16 conditions of the kind «the setup happened in the last 3, 5 or 10 bars, not counting this one»; the other condition of the strategy is the trigger, on the current bar:
- An event: Count(BullDivRSI(14) = 1, 5)[1] >= 1 (a divergence, a chart pattern, a candlestick pattern)
- An oscillator crossing a threshold: Count(RSI(Close, 14) crosses below 30, 10)[1] >= 1
- The close crossing a level: Count(Close crosses above BB_Upper(Close, 20), 3)[1] >= 1
This way a divergence or a double bottom, which on the exact bar hardly ever coincide with anything else, keep the entry armed for a few bars. The [1] makes the setup have to come before the trigger. You can write the same by hand in the AQL editor, for example Close crosses above SMA(Close, 20) and Count(RSI(Close, 2) < 10, 5)[1] >= 1. They are ordinary formulas: GPU, fingerprint and all six platforms.
Indicators against their own past: the shift [n]
Genesis can compare an indicator with its value N bars ago, opening up classic trend and momentum signals that would otherwise be impossible:
A rising average
Momentum
RSI accelerating
It is controlled in ⚙️ Settings → The search, with Shift [n] (From / To / Step). From = To = 0 disables it; a small step generates far more combinations (slower).
Flat out: GPU + speed control
Genesis makes full use of your hardware: the genetic algorithm runs on the GPU (OpenCL) and the Survival test is shared across all the CPU cores — it researches much faster, and whatever enters the Room is still re-validated on CPU (parity intact). The ⏱️ Performance profile panel shows you where the time goes, phase by phase.
You decide the speed / rigor balance on each run:
- → Robustness markets (in ⚙️ Settings): on how many markets each strategy is re-tested. None = maximum speed; a handful = a balance.
- → Write the log: if you do not need that run's Viewer/CSV, turn it off to go faster still.
- → Log to disk (next to Record log, unticked by default): saves a copy of the log to a text file (Genesis\bitacoras folder) with the run's settings at the top, to review it calmly afterwards.
- → Unticking a rung (e.g. Survival) skips that test entirely.
Why Genesis flies: smart caches
Researching non-stop, Genesis was repeating a lot of identical work over and over. The idea is simple: do it once and reuse it — like a cook who buys the ingredients once for the whole service instead of going to the market for every dish. Three caches gave it a ~3× speed-up:
Each robustness market is loaded once per run, not once per strategy.
Each group's "alphabet" of conditions is built once, not on every round.
The data is uploaded to the card once (the full pool) and the rounds reuse it.
You notice it in two places: the ⏱️ Performance profile breaks down what it spends the time on (uploading to the GPU, computing on the GPU, building the pool and the population, evolving, Survival…) so you can see at a glance where the bottleneck is; and the Last decision panel shows the genetic algorithm's progress live («Generation 15/40 · best fitness…»). None of this changes the results: whatever enters the Room is still re-validated on CPU, identically.
AniQuant can be tried free for 14 days, with every module and no card. If you have the book, 30.