Sep 22
StrategyQuant X Build 145: A Deep Look at What Is Coming
Build 145 of StrategyQuant X comes out in about a month, some time in October. I’ve been working in the release candidates for several weeks now, and the latest one, rc3, came out on 17 September. This post goes through what’s in it, feature by feature, for people who already use StrategyQuant and want to know what they’ll get.
I’m StrategyQuant’s AI Solutions Architect and some of this is my own work, so read it as an inside preview. StrategyQuant links are affiliate links. Names and numbers come from rc3 and may still change before release. When it’s out, it will be on the download page and the What’s New page.
There are four big pieces:
- The AI assistant now does the work itself. It runs projects, reads databanks, backtests and checks its results.
- A Marketplace tab where you install add-ons with one click and publish your own.
- 69 new blocks built on Commitment of Traders positioning data.
- Volume Profile grows from 4 signal blocks to 36, with ready templates.
There are smaller changes too: Anchored VWAP, Java 25, and Python with pandas and matplotlib bundled in.
1. The AI assistant now does the work itself
In Build 144 you typed a request and got an answer back. That was it.
In 145 a request starts a loop. The assistant plans a step, calls a tool, looks at what came back and decides what to do next. Then it repeats that until the job is done. So one message like “build a breakout project on EURUSD H1 and check the survivors out of sample” turns into a few dozen actions. It sets up the tasks, starts the run, waits for it, reads the databank and retests the best strategies on out-of-sample data. If a setting was wrong, it fixes it and runs again.
It works on your real platform. It reads your actual databanks, changes real project settings, starts and watches runs, backtests a saved .sqx when you ask, and writes and compiles custom code. The numbers it reports come from those reads.
When two tool calls don’t depend on each other, they run at the same time. Reading five strategies takes one round.
Helper agents
The main assistant can hand parts of a job to smaller specialist agents. Each has its own instructions and usually a short list of tools it’s allowed to use. Several can run in parallel while the main one waits and puts their answers together.
My favourite is research-critic. Its only job is to try to prove a finding wrong before it gets written down. It can read databanks, stats, strategy code and files, and that’s all. It can’t run anything or change anything.
I like this setup a lot. When you ask a model to review its own reasoning, it will almost always approve of it. A second model that’s there to find the flaw usually finds it. I wrote about this in why an agent that grades its own homework tells you nothing, and it’s good to see the split built into the product instead of left to whoever writes the prompt.
Research that runs overnight
A turn can end with the assistant scheduling its own next step. SQX wakes it up later and it carries on, even with the panel closed. That makes an overnight research loop possible. It forms a hypothesis, builds, tests out of sample and lets the critic attack the result. Then it writes the verdict into the journal and moves on to the next idea. You can always see a scheduled wake-up and cancel it with one click.
You set how far it can go: a maximum number of tool rounds per turn and a time budget for each tool. Without those limits I wouldn’t leave it running overnight.
Memory
The assistant keeps its memory in plain Markdown files that you can open in any editor:
| Layer | What goes in it |
|---|---|
| Long-term memory | Short facts about you, like “I trade only EURUSD and GBPUSD” or “judge by Return/DD, not net profit”. Loaded into every conversation. |
| Daily fact ledger | Written automatically after each turn and merged into long-term memory in the background. |
| Knowledge base | One entry per finished piece of research with its verdict, a profile of you as a trader, and a list of open threads. |
| Learned skills | When a procedure works well, the assistant saves it as a skill for itself. Each one is security-scanned before it loads. |
/recall EURUSD breakout gives you what you already know about a topic: what was tried and the number that decided it, what was rejected, what’s still open, or that there’s nothing yet. /journal records a verdict. When a result was inconclusive, it’s recorded as inconclusive. I’ve written about deciding what counts as a pass before you test, and this journal gives that habit a home inside the platform.
Watching it work
Everything shows up in the panel live: each tool call, each helper-agent step with its name, and a timer for the turn. The Stop button ends a turn cleanly.
There are a few safety features, and I think they matter more than the flashy ones:
- File access is limited to the SQX data folder. To open another folder, type
/allow-path D:\my\folder, and addroif you want it read-only. The permission lasts until you start a new chat. - Every change is snapshotted. An operation log records what changed, and a restore tool can undo it. If you ask “what have you changed in this project so far?”, the answer comes from the log.
- The assistant can always see the screen and move around the app. Clicking controls stays locked until you type
/drive-ui, and it can’t unlock that for itself. - Shell commands go into an audit log. Skills the assistant writes for itself are scanned and can wait for your approval before they load.
By default it runs on StrategyQuant’s own models and uses AI credits, and each answer shows what it cost. If you’d rather pay the provider directly, you can add your own API key for OpenAI, Anthropic, Google or a local model server.
2. AI plugins
Everything the assistant knows how to do comes from AI plugins. Each plugin is a plain folder under user/extend/AIPlugins/ with a plugin.json file and some Markdown files. There are four kinds of item inside:
| Kind | What it is | When it’s used |
|---|---|---|
rules/ | Always-on instructions, like “never risk more than 1%” or “report in a table” | Added to every conversation |
skills/ | A step-by-step procedure for one type of task | Loads automatically when the task matches, or when you type /name |
commands/ | A saved prompt with a $ARGUMENTS placeholder | When you type /name args |
agents/ | A helper agent with its own prompt and optional tool list | Called automatically, or with /name task; several can run at once |
Put a folder in there and it works from your next message. You don’t need to restart or register anything. Type / in the chat box and you’ll see a list of everything your installation can do.
Each item has a runtime field that says where it runs. sqai means inside the app only, claude-code means the terminal only, and both is the default. Most items are both. The app-only ones need live project state.
Build 145 ships 17 plugins with 36 skills, 11 slash commands, 11 helper agents and 13 rules. These are the ones I use:
| Plugin | What you use it for |
|---|---|
| strategy-analyst | Explains a strategy’s logic in plain language, judges its results and suggests improvements. Includes /deep-audit and /strategy-report. |
| strategy-architect | Sets up a Builder project for a style (mean reversion, trend, breakout, martingale) or a full multi-task custom project. A config-validator agent checks the setup before anything is applied, and you have to confirm first. |
| quant-research | Five research workflows. /robustness-screener groups a databank’s winners by trading idea, so you notice when ten “different” strategies are really one idea. Also /portfolio-precheck, /databank-report, /config-ab-test (snapshot, run A, run B, compare, restore) and /morning-digest. |
| auto-research | The overnight loop, with research-critic inside and an oos-discipline rule that keeps the holdout data untouched. |
| trading-memory | The knowledge base, /recall and /journal. |
| market-analyst | Classifies regimes (trending or mean-reverting, volatility, session and day-of-week effects) from exported bars, and breaks down trade lists (expectancy, MAE/MFE, streaks). The numbers come from numpy code, so they’re the same every run. |
| sqx-snippets | Eleven skills for writing SQX snippets: indicators, signals, databank columns, custom stats, money management, Monte Carlo methods, What-If methods, order actions and trade analysis. |
| sqx-reports | Report recipes and a Python library. Your data stays in files on your machine, and the model only gets a short JSON summary. |
| analysis-tabs | Builds HTML plugins that show up as extra tabs in the Results panel. |
| sqx-strategy, ac-strategy, sqx-project, algowizard-lab | My own tools from sqx-lab and the agentic framework, rebuilt to run inside the platform. They turn a sentence into a verified .sqx for the standard engines or the AlgoCloud stockpicker engines, build a whole custom project from one spec, and create custom blocks and block groups. |
| plugin-maker | /make-plugin turns a workflow, checklist or trading protocol you paste in into a proper plugin. |
| ui-pilot | Drives the interface where you can see it, after you type /drive-ui. |
| breakout-project-factory | Clones breakout project pipelines and checks them, using one agent to build and a separate read-only agent to verify. |
| custom-page | New in this build. Add your own HTML page to the main menu with just a module.js and an index.htm. No Java needed. |
The part I like most is that you can extend the assistant with plain text. If you keep a research checklist that you paste into every chat, /make-plugin turns it into rules, a skill, agents and a command. After that it’s there in every new session without pasting.
3. Marketplace
There’s a new shopping-bag icon in the navigation, and it opens the Marketplace.
Installing takes one click, and you can watch it go through four steps: downloading, verifying, installing, done. The package is downloaded and its SHA-256 checksum is checked. Then it’s unpacked into your user folder, and every file it added goes into a small local database. Because of that, uninstalling removes exactly those files and nothing else. Packages describe themselves in a manifest.json and install into your extend/ and custom_indicators/ folders.
Publishing happens in a form inside the app. You fill in a title, a short description and a longer one with a proper text editor. Then you pick a main category and extra ones and enter a version number. The author is filled in from your licence. You add a cover image and up to five gallery images, and choose files from a tree of your own shareable files. Your settings, customdata and cache folders don’t appear in that tree, so you can’t share them by accident. Packages are zipped with a new .sqm extension. StrategyQuant can turn on review, and when it’s on, your submission waits for approval before it goes public.
This one matters to me personally. I’ve published 135 custom indicators and tools in the Codebase for free, and until now every one of them had to be installed by hand. You download a zip, figure out which folder it goes in, unpack it, restart and hope it works. I’ve answered “where do I put this file” by email more often than I’ve written indicators. Clicking a button to install, with a checksum check and a clean uninstall, is a big step up.
4. COT: trading on who holds what
This is the new part that surprised me most, because it lets you ask a different kind of question.
Nearly every indicator you’ve used is calculated from price. Moving averages, oscillators, breakouts, volatility bands all start from the same price data and rearrange it.
The Commitment of Traders report is different. The CFTC publishes it every week, and it shows what different groups in the futures market actually hold. Commercial hedgers use the contracts for their business. Large speculators are there to follow trends. Small traders tend to arrive late. It tells you about positioning, which price data doesn’t.
Build 145 adds a new COT category with 69 blocks: 11 numeric indicators and 58 true/false signals.
The indicators
A raw count of net contracts doesn’t tell you much by itself. So each group’s net position is compared to its own range over a rolling window and scaled from 0 to 100. The default window is 156 weeks, about three years, and 52 weeks is also available. A reading of 95 means that group’s positioning is near the top of its three-year range, and you can build a rule on that.
From that base you get:
- positioning indexes for commercial hedgers, large speculators and small traders
- a composite score that combines them
- hedger-versus-speculator spreads and cross-field spreads
- momentum on the score
- a block that reads a raw field value
- a block for building your own combination of fields
The signals
The 58 signals are ready-made versions of ideas you’d otherwise code by hand:
| Family | Examples |
|---|---|
| Extremes | Bullish and bearish extreme, extreme exit, extreme persistence, extreme unwind, speculator exhaustion long and short |
| Crowding | Crowding unwind, smart money versus dumb money, small-trader fade |
| Divergence | Higher low and lower high against positioning, triple positioning divergence, hedger divergence confirmation |
| Regime shift | Regime shift up and down, 52-week versus 156-week shift, seasonal |
| Momentum | Hedger acceleration, score acceleration, momentum reversal, rate of change, score above and below, score zero cross |
| Cycle and exits | Full-cycle long and short, full-cycle exits, emergency exits for longs and shorts |
| Statistical | Spread z-score, neutral zone, convergence, accumulation, multi-window agreement |
| Other | VIX proxy pair, speculator trend long and short |
How it fits your charts
COT data comes out once a week and is published a few days after the positions are recorded. On every bar of your strategy, the blocks use the latest report that was already public at that time, and keep using it until the next one appears. A daily or H1 strategy can’t see positioning before it was published. If you do this by hand and get it wrong, the backtest looks great for reasons that have nothing to do with trading. Momentum and lookback settings on these blocks count weekly reports, not chart bars, so keep that in mind when you set them.
The data comes from SpreadCharts. You get it into SQX by running a downloader once: an EA for MetaTrader 4 and 5, or a NinjaScript version for NinjaTrader. It builds a local cache. Your chart symbol is matched to the right futures contract automatically from a table of more than a hundred symbols. Check that your symbol points to the contract you expect. And because the default window is 156 weeks, load several years of history.
The blocks export to MetaTrader 4, MetaTrader 5, NinjaTrader 8, EasyLanguage (TradeStation and MultiCharts) and pseudocode.
The guides
Two Word documents come in the COT/ folder. The first is a block reference with a quick start, notes on which markets suit COT, and an appendix on scaling, timing and data alignment. The second is a strategies guide with 39 numbered strategy designs, running from “Hedger vs Speculator Extreme Reversal” and “Crowding Unwind” through “52w vs 156w Positioning Shift” to “COT Full-Cycle Positioning”. It also covers COT basics, research findings, using COT as a machine-learning feature, and a do and don’t checklist.
Both guides give the same advice, and I agree with it. COT tells you the background positioning, and price tells you when to trade. A crowded position can stay crowded for months. Use COT as a filter for the market regime, the same way I use the regime detection indicators I wrote about earlier, and let price give the entry.
5. Volume Profile, the full set
Volume Profile and TPO arrived in Build 144 with four signal blocks. That was enough to look at but not enough to build strategies with. Build 145 has three indicators and 36 signal blocks in nine categories.
This is my colleague Emmanuel Evrard’s work. The blocks cover what a volume profile can tell you in an organised way, and its documentation is the best of any block family in the platform.
The idea behind it is simple. A moving average shows where price has been. A volume profile shows where trading actually happened, meaning where the orders were and where they’re likely to show up again. It covers the point of control (POC), the value area, the session’s initial balance, and the thin areas where hardly anyone traded.
The three indicators are:
- the session Volume Profile
- a custom-hours version
- a multi-session version with four sessions you can set (London, New York, Sydney, Tokyo)
All three give you the point of control, value area high and low, initial balance, high and low volume nodes, delta, and separate bull and bear POCs.
The signals by category:
| Category | What it checks |
|---|---|
| Price versus structure | Price above or below the POC, inside or outside the value area |
| Order flow and delta | Bull and bear POC relationships, delta thresholds |
| Profile shape | Where the POC sits in the profile, long and squat shapes |
| Value area width | Narrow and wide value areas |
| Initial balance | Breakouts above and below, and rejections of both |
| POC dynamics | POC rising or falling from session to session |
| Value area dynamics | Value moving up or down from session to session |
| Reversal patterns | Excess reversal long and short, open drive long and short |
| VWAP | Cross up and cross down |
It also comes with:
- Ten ready
.sqxtemplates: daily session versions, a swing-session version, and H1 templates built on the POC and value-area movement signals. - A random group with every VP signal, so the Builder can search the whole set without you wiring each block in.
- A Custom Project preset.
- A separate TPO folder with its own build config and an MT5 template based on TPO market profile.
Emmanuel’s documentation (version 2.0, dated 14 July 2026) explains the concepts, the three indicators, all nine signal categories and the common parameters. It includes four example combinations, for trend-following, mean-reversion, failed-breakout and distribution setups, and a table of recommended timeframes and sessions. For the plain profile indicator, it suggests H1 with a monthly session or M30 with a weekly one.
6. Also new
Anchored VWAP grows to 13 blocks. There’s the anchored VWAP itself, with standard-deviation bands that reset at the start of each day, week or month. Then there are blocks for rising and falling, for closes above and below the line and each band, four band-cross variants, and a fast versus slow comparison.
Java 25 is now under the hood.
Python 3.13 is bundled, separate from any Python you already have, with numpy, pandas and matplotlib. The assistant uses it to work out things the Results tabs don’t show, like return correlations, day-of-week and hour-of-day effects, volatility regimes and custom robustness metrics. If you’ve read how I use Python alongside StrategyQuant, this is the same workflow without the setup.
AI Reports. Bigger results can become a standalone report page with KPIs, tables and charts. It’s shown in a sandbox with no access to the app, your files or the internet, and it’s still there after a restart.
Attachments in the chat. Paste a screenshot of an equity curve or drop in a trades.csv and ask for a drawdown breakdown.
If you’re still deciding whether StrategyQuant is right for you, my review covers that. This build adds a lot to the part of that review about extending the platform.
Frequently asked questions
What is new in StrategyQuant X Build 145?
Four big things. The AI assistant now works as an agent: it plans a step, runs tools, reads the results and keeps going until the job is done, hands parts of the work to specialist helper agents, and can schedule itself to continue research overnight. A Marketplace tab arrives with one-click, checksum-verified installs and an in-app form for publishing your own work. A new COT category adds 69 Commitment of Traders blocks built on weekly futures positioning data. Volume Profile grows from 4 signal blocks to 36, with three indicators, ten ready strategy templates and a random group for the Builder. Anchored VWAP grows to 13 blocks and Java moves to version 25.
How does the AI assistant work in StrategyQuant X 145?
Every request runs as a loop. The model plans a step, calls one or more tools, reads what came back and continues, round after round, within limits you can set for tool rounds per turn and time per tool. Tool calls that do not depend on each other run in parallel. Helper agents with their own instructions and narrow toolsets take sub-tasks and run at the same time. A turn can end by scheduling its own wake-up, which is how unattended overnight research works. Memory is kept as plain Markdown files, and the assistant can save a procedure that worked as a new skill for itself.
What is the StrategyQuant Marketplace?
A new tab in the navigation for installing and sharing add-ons. An install downloads the package, checks its SHA-256 checksum, unpacks it into your user folder and records every file it added in a local database, so uninstalling removes exactly what was installed. Publishing is a form inside the app: title, descriptions, categories, version, author taken from your licence, a cover image, up to five gallery images, and a file-tree picker over your own extend folder. Packages use a new .sqm extension, and submissions can go through review before they are published.
How many COT blocks are in StrategyQuant X 145?
69 in a new COT category: 11 numeric indicators and 58 true or false signals. The indicators are positioning indexes for commercial hedgers, large speculators and small traders, scaled 0 to 100 inside a rolling window, plus a composite score, spreads, momentum and a block for building your own combination. The signals cover extremes, crowding, divergences, regime shifts between the 52-week and 156-week view, momentum, z-scores, full-cycle and emergency exits, and a VIX proxy pair.
Does the COT data need a separate download in StrategyQuant?
Yes. The weekly positioning data comes from SpreadCharts and reaches the platform through a downloader you run once per platform: an EA for MetaTrader 4 and 5, and a NinjaScript version for NinjaTrader. Your chart symbol is mapped to the right futures series automatically through an alias table covering more than a hundred symbols. The default window is 156 weeks, so load several years of history.
How many Volume Profile blocks are in Build 145?
Three indicators and 36 signal blocks in nine categories. The indicators are the session Volume Profile, a custom-hours version and a multi-session version with London, New York, Sydney and Tokyo sessions, all giving point of control, value area high and low, initial balance, high and low volume nodes and delta. Ten ready .sqx templates, a random group for the Builder, a Custom Project preset and a TPO build config come with them.
When will StrategyQuant X Build 145 be released?
In about a month, so some time in October 2026. It is at release-candidate stage now; my install is rc3, build 3148, from 17 September 2026. The final build will appear on the download page and the What's New page. Details in this article come from the release candidate and can still change.