> ## Documentation Index
> Fetch the complete documentation index at: https://docs.quantspace.limex.pro/llms.txt
> Use this file to discover all available pages before exploring further.

# Pipeline

> How a QuantLab research flow usually unfolds

QuantLab is not a single fixed recipe. Depending on what you ask for, the agent chains different steps. Think of it as a few common **paths** rather than one mandatory sequence.

For dates: without ML you usually name only the sample **start** and **end**. The part you evaluate is that whole window except the first `covariance_window` days used as warmup. With ML/DL you still think in start + end of the sample — not a three-ended train / mid / test calendar.

***

## Common paths

**Quick backtest** — you care about weights and historical performance, not machine learning:

```
market data → portfolio construction → backtest
```

**Forecast-driven research** — you want models to estimate returns before building the book:

```
market data → features → ML or deep learning → portfolio construction → backtest
```

**Stress and “what if”** — you already have a portfolio (or just tickers and weights) and want scenario losses:

```
market data (and your weights) → stress tests / risk views
```

Often that sits **on top of** a backtest path: after you have a constructed portfolio, you ask for charts, tearsheets, factor stories, or multi-model stress in the same conversation.

You do not have to name these steps yourself. Describe the goal; the agent picks the sequence.

***

## Mixing and matching

A few patterns people use often:

* **Same data, many portfolios** — download prices once, then try equal weight, HRP, max Sharpe, DCC, and so on side by side.
* **Backtest, then dig in** — after a run, ask for an equity curve, weight paths, or a risk report without starting from scratch.
* **Your book, QuantLab analytics** — paste holdings and ask how bad tomorrow looks under a shock, or whether something looks overweight versus optimized alternatives.
* **With or without ML** — skip models when you only need classical portfolio math; add XGBoost, LightGBM, CatBoost, or a neural net when forecasts matter.

Independent asks (several optimizers, or charts + stress + risk after one portfolio) can run **in parallel**. Steps that truly depend on each other still wait their turn.

***

## Keeping results

Each finished step leaves a result link you can reuse later — so you can change one assumption, compare methods, or come back tomorrow without re-downloading everything. [Example prompts](/best-practices) show how that sounds in chat.