[ getting started · 3 of 3 ]

Your first backtest

Run a walk-forward backtest of a built-in sleeve, watch cuFOLIO solve at 520 ms a step, and save the winner as a loadable strategy. This is the fastest path to seeing GB10's optimization throughput on real data.

Open the Backtesting page

From the top nav, click Backtesting. You land on the run-builder. The left panel is the run config; the right panel is the live equity curve once a run is in flight.

Pick a preset

Click + New run.
Choose preset Momentum top-decile (a balanced sleeve; 8 symbols, monthly rebalance, CVaR α=0.95).
Leave the universe as SPY · QQQ · NVDA · MSFT · AAPL · META · AMZN · GOOGL.
Date range: last 18 months.

The right side shows estimated wall-clock based on the universe size and rebalance count. For this preset, expect ~12 seconds end-to-end.

Solve once (no backtest)

Before the full walk-forward, click Solve once (no backtest) →. This fires a single cuFOLIO solve and shows you the telemetry tiles:

scenario gen   269 ms  ← KDE on GPU
LP solve       251 ms  ← cuOpt PDLP
total          521 ms
α              0.950
scenarios      5,000
constraints    long-only, max_pos 25%, cash >= 2%

The weight bar chart shows the optimizer's allocation under the current preset. Take a screenshot — this is the cuFOLIO GPU solve telemetry.

Run the walk-forward

Now click Run backtest →. Watch the equity curve update as each rebalance step completes. The history table at the bottom appends a row when the run finishes:

metric	typical value	what it means
CAGR	+24.7%	annualized compound return
Vol	16.4%	realized annualized vol
Sharpe	1.51	(CAGR − rf) / vol
Max DD	−9.8%	worst drawdown over the window
Turnover	62%/yr	sum of `\|Δw\|` per rebalance, annualized

AutoResearch sweep

Toggle AutoResearch loop on the run config, then click Run →. NVTrader sweeps over a small grid of (α, max_position_pct, n_scenarios) candidates, runs a full walk-forward for each, and surfaces the winner card with one click to save.

The sweep typically runs 8 candidates in about 30 seconds on GB10. The winner's saved-strategy file lands at configs/sleeves/auto_<timestamp>.yaml; it loads instantly next time.

Save as a strategy

Click Save as strategy on the run row. Name it; pick a sleeve (Core/Sat/Tac). The strategy is now selectable on the Portfolio page's rebalance modal and shows up in configs/sleeves/user_*.yaml.

Next steps

Tune cuFOLIO — adjust α, max position, or constraint behavior.
Trigger an end-to-end cascade from Observability and watch the 21-agent system run.
Train a NemoRL policy on the same universe and compare against the cuFOLIO proposal.