Roadmap Inquiry

[JungleWolfz]

Is there a plan to support minute-level factor-enhanced medium- and low-frequency strategies? See the article below: https://zhuanlan.zhihu.com/p/2030576049969419663

[Miasyster]

Thanks for the interest and the great reference! The DolphinDB article describes a solid approach — aggregating minute/tick-level microstructure data into daily factors for medium/low-frequency strategies.

Current status: QuantGPT currently operates at daily frequency only (OHLCV + fundamentals). This is explicitly listed in our Limitations section.

Our thinking on this direction:

The "high-frequency data, low-frequency signal" approach is valuable — factors like smart money flow, realized jump volatility, and intraday VWAP deviation carry real alpha that daily bars can't capture. We see two possible paths:

1. Intraday-derived daily factors (lower barrier): Pre-compute minute-level features into daily factor columns (e.g., smart_money_score, intraday_reversal), then plug them into the existing expression parser and backtest engine. This preserves the current architecture while enriching the factor space.

2. Native minute-level backtesting (higher barrier): Support minute bars directly in the backtest engine. This requires a different data pipeline, much larger storage, and rethinking the expression parser's time-series operators — a significant architectural change.

Path 1 is realistic for a near-term iteration. Path 2 would be a major version milestone.

The bottleneck is data, not code. Free data sources (baostock, akshare) don't provide reliable minute-level history at scale. Professional sources (rqdatac, Wind, DolphinDB) require paid subscriptions. We'd need to either integrate a paid provider or support user-supplied data.

Keeping this issue open to track interest. If you (or others) have specific use cases or preferred data sources in mind, that context would help prioritize.

[JungleWolfz]

data sources ：
1.https://oss.quantmindai.cn/qmt-agent.html，qmt is a data source provided by brokerages (such as Guojin and Guosheng, with Guojin being the most common), and the link is an example.
2.https://github.com/LisonEvf/opentdx Tdx data is more stable than sources such as akshare, and it also provides minute-level data.
Native minute-level backtesting ：
wtpy quantaxis akquant

[Miasyster]

Thanks for the concrete suggestions @JungleWolfz!

I looked into each of them:

Data sources:

• QMT (XtQuant): Great data coverage (minute/tick), but requires a brokerage account with minimum deposit (¥100K+ for most brokers) and only runs on Windows. Not viable as a default data source for an open-source project where users should be able to get started without a specific brokerage setup.
• OpenTDX: Lightweight and clever (reverse-engineered TDX protocol, no client needed), but the author explicitly marks it as a learning project, not production-ready. Minute-bar history depth is limited (~800 bars per request), which isn't enough for multi-year backtests. The reverse-engineering approach also carries stability and legal risks.

Backtest frameworks:

• wtpy: Solid C++ core but largely inactive since late 2023.
• QUANTAXIS: Large ecosystem (10K+ stars) but heavy legacy baggage, core maintainer inactive.
• akquant: The most interesting one — genuine Rust+Python (maturin/PyO3), active development, MIT license. However, it's an event-driven strategy backtester (order/fill/settlement), not a factor research tool. It lacks IC, quantile-group analysis, turnover metrics, and factor scoring — which is QuantGPT's core. Two useful engineering ideas worth learning from: (1) AST-based automatic materialization for nested TS/CS operator conflicts, (2) golden baseline testing with Parquet checksums.

Our current thinking remains:

The bottleneck for minute-level support is still data availability from free sources. Path 1 (pre-aggregate intraday features into daily factors like smart_money_score, intraday_momentum) is the pragmatic near-term approach — it enriches the factor space without requiring architectural changes or paid data subscriptions.

That said, if users bring their own minute-bar data (from QMT, TDX, or other sources), we could support a "load custom data" workflow. That might be a reasonable middle ground.

Appreciate the research pointers — keeping this tracked here.

[zaleqzzhang]

Supporting minute-level data represents a significant capability enhancement for the current project. While it is true that data sourcing can indeed be a barrier for individual open-source projects, this should never be the sole justification for removing an important feature. At the framework level, it may be acceptable not to provide a specific data source solution; however, it should at least offer a configurable option in terms of capability. Such proactive support from the framework level would also make the system more acceptable to users who have the ability to address the data sourcing challenge themselves.

[Miasyster]

Good point — and this aligns with what I mentioned above about the "load custom data" workflow. To be clear: the plan was never to skip minute-level capability entirely because of data sourcing. The intent is:

1. Framework-level: expose a generic data adapter interface so users can plug in their own minute bars (from QMT, TDX, Wind, or any source). The backtest engine and expression parser would handle multi-frequency data through this interface.

2. Default data: continue shipping with free daily sources (baostock/akshare) out of the box, so new users can get started with zero setup.

In practice, this means a user with QMT access could load their minute data and run intraday-derived factor research, while someone without it still gets the full daily workflow. Framework capability and data sourcing are treated as separate concerns — which is exactly what you're suggesting.

I'll track the data adapter interface as a concrete milestone. Appreciate the push to make this explicit.

[Miasyster]

Update: Data Adapter Interface is Live

We've built and open-sourced the generic data infrastructure layer we discussed here:

👉 adata — Agent-Native financial data infrastructure for A-shares.

This directly addresses @zaleqzzhang's point about separating framework capability from data sourcing.

What adata provides

• BaseProvider abstract interface — implement fetch_daily, fetch_benchmark, list_instruments, list_universe and plug in any data source
• 5 built-in providers: baostock, akshare, tushare, rqdatac, PolarDB — with automatic fallback chain
• Parquet cache layer — cache-first, incremental merge, zero-config local storage
• MCP server — any LLM agent can discover and query data via standard MCP protocol
• Code normalization — handles sh.600519 ↔ 600519.XSHG ↔ sh600519 transparently

What's next

Minute-level support is the natural next step. The architecture is ready — it needs a fetch_minute method on BaseProvider and corresponding store logic. Users with QMT/TDX access could implement their own minute-level provider today by subclassing BaseProvider.

QuantGPT will integrate adata as its data layer in a future version, replacing the current built-in market_data module.

cc @JungleWolfz @zaleqzzhang