Backtest

Native replacement for legacy backtest page. Runs walk-forward backtest on one ticker at a time.

Company / Ticker

Years

Step Days

Eval Days

Horizon

Ready. Enter a ticker, then click Run Backtest.

⚙️ Signal Tuning

BUY Threshold 0.10

SELL Threshold -0.10

Min Agreement 60%

Min R:R Floor 0.00

Use Time Stop On

📊 Portfolio Overlay (Universe Backtests)

Corr Threshold 0.75

Max Corr Weight 12%

Target Vol % 2.5

Trade List (Most Recent Window)

Date	Action	Entry	Exit	Return%	Hit	Score

How to interpret this: Compare AI return vs buy-and-hold first, then check hit rate and profit factor. A high hit rate with weak total return often means winners are too small vs losers. For unstable periods, shorten step/eval windows and compare consistency across horizons.

How To Use This Backtest

Start with a 5-year run on your ticker, then compare a second run with slightly tighter or looser thresholds. Prioritize Alpha vs B&H, Profit Factor, and Max Drawdown over hit rate alone.

Total Return vs Buy/Hold

If AI return is positive but Alpha vs B&H is very negative, the model is profitable but not beating passive trend capture.

Hit Rate

Useful, but not enough alone. A 60% hit rate can still lose money if losers are much larger than winners.

Profit Factor

Gross wins / gross losses. >1.5 is solid, >2.0 is strong, near 1.0 means weak edge.

Sharpe Ratio

Risk-adjusted return. Around 1 is acceptable, >1.5 is strong, <0.5 often means unstable performance.

Target Hit % / Stop Hit %

High stop-hit with low target-hit usually means entries are early or thresholds are too loose for that regime.

RL Steps

Number of learning updates performed during this backtest run. Higher steps means more training signal for the agent.

Tuning playbook:

If Alpha vs B&H is deeply negative in bull periods: lower BUY threshold slightly or reduce Min Agreement to allow more participation.
If drawdowns and stop hits are high: raise Min Agreement and set a positive Min R:R floor (for example 0.8 to 1.2).
If win rate is high but return is weak: improve payoff ratio, not win rate. Focus on Avg Win/Loss and Profit Factor.
Run at least two horizons (Short-Term and Mid-Term) before deciding a ticker has no edge.

Best practices to increase learning quality:

Use 5-year history when possible and test at least two horizons so the agent sees multiple market regimes.
Avoid one-and-done runs. Use 3-run cycles per ticker (baseline, stricter filters, balanced filters), then keep the best risk-adjusted profile.
Keep output windows realistic: step_days 20-30 and eval_days 10-20 are usually stable for web backtests.
Do not optimize for hit rate alone. Prefer settings that improve Sharpe and Profit Factor while controlling drawdown.
Retrain across diverse sectors (not only mega-cap tech) so learned behavior generalizes better.

How scans are affected by learning:

Universe scans and single-ticker analysis generate actionable signals that are logged for outcome evaluation.
When hold windows mature, outcomes are evaluated in the background and fed back into RL as rewards.
As this loop accumulates, scan output can shift: confidence and action mix may change as the policy adapts to what has recently worked.
If outcomes are still pending, immediate scan behavior may look similar. Meaningful changes appear after enough evaluated outcomes accumulate.