How Quants Work
The people, process and discipline behind a professional systematic shop — and where QSYS maps to it
A quant shop is an assembly line, not a genius with a formula. Ideas flow through a gated funnel and the majority are killed on purpose — that kill rate is the quality control. The durable edge is the enforced process: rules for every idea, every size, and every kill, plus the compounding of many small, weakly-correlated edges. This page walks through how that process works and how the tools in QSYS fit into it.
1 · The team and how it interacts
Five roles, each owning one stage of the line.
Quant Researcher
Generates and tests signals. Most of the job is trying to disprove an idea — most patterns are noise.
Data Engineer
Wrangles and cleans feeds, proving the data doesn't leak future information (point-in-time, survivorship-free).
Quant Developer
Builds the platform: data pipelines, the backtest engine, execution, monitoring.
Risk Manager
Sets position limits, exposure caps and drawdown thresholds — and holds the authority to cut positions without asking.
Portfolio Manager
Owns the book, allocates capital across strategies and, in pod shops, gets cut after losing a few percent.
2 · The strategy lifecycle
A gated funnel. Single-digit percentages of ideas ever reach capital.
- 1
Testable hypothesis — a mechanism first, not a bare data-mined correlation.
- 2
Data acquisition & cleaning — often the longest step.
- 3
Backtest with realistic costs and slippage.
- 4
Robustness checks — does the edge survive across time, assets and parameters?
- 5
Walk-forward / out-of-sample on data never touched during tuning.
- 6
Paper trading on live data.
- 7
Small live allocation.
- 8
Scale — or kill. Survivors routinely die at OOS because the researcher unknowingly tuned to the test set.
QSYS maps here: the Walkforward tool is step 5 — out-of-sample validation with rolling re-selection.
3 · Backtest hygiene — not fooling yourself
An edge that's brilliant gross and flat net is the most common species in quant research.
Costs & slippage
Model commissions, bid-ask spread, slippage and market impact. Model Capacity — how much capital the edge can actually absorb.
Point-in-time & survivorship
Use data as it existed at the time. Include delisted names. Never use revised numbers that weren't public yet (look-ahead bias).
Overfitting / multiple testing
Every parameter grid-search inflates Sharpe. Adjust for the number of trials with the Deflated Sharpe Ratio or the Probability of Backtest Overfitting. Look for a plateau of profitable parameters, not a single spike — if a ±10% parameter change kills the strategy, it's overfit.
The Dimensional study
Of 720 market-timing strategies, 30 "worked" and 690 didn't — most by pure chance. Out-of-sample validation is not optional.
4 · Risk management in production
Sizing is a risk preference, not an optimizable parameter.
Volatility targeting
Size inversely to realized vol; scale down when vol spikes. Built-in risk control in managed futures.
Fractional Kelly
Half- or quarter-Kelly as a ceiling, never full Kelly — you never know the true edge.
Layered risk limits
Pre-trade limits, in-flight throttles, and a kill-switch independent of the desk.
Drawdown & factor caps
Cap factor exposure, not just raw vol, so no single concentrated sector dominates.
QSYS maps here: the drawdown circuit breaker, the regime-based risk guard, and crisis replay in the Stress Test.
5 · Monitoring a live strategy
A live strategy isn't "done"; it drifts. Track the live-vs-backtest gap in realized slippage and fills. Do performance attribution by source — it's the only research input not contaminated by hindsight. Kill when live results stop tracking the backtest — within months, per pod-shop discipline.
QSYS maps here: PnL Attribution and the Run Registry (every backtest pinned to its config, git commit and metrics).
6 · Culture & reproducibility
Research lives in Python/R notebooks, but professional shops enforce reproducibility: version-controlled code, configs and data, pinned environments, and a run registry linking each backtest to a git commit, dataset version, params and seed. Review happens before anything is promoted to capital. That enforced process — not any single formula — is the edge.
This is exactly what the Run Registry implements on QSYS.
Why this matters
Most individual investors optimize a single backtest. Professional quants optimize a process: many small independent edges, honest costs, strict out-of-sample discipline, layered risk controls, and a kill rule for anything that stops working. The tools in QSYS exist so you can run the same discipline on your own book.