Research dashboard / crypto strategies
Out-of-Sample Strategy Trade Charts (Crypto)
A visual review of 250 adaptive crypto trades sampled across the return distribution.
Charts use research-database OHLCV data with asset_class = crypto,
resampled to 1-hour candles, with the exported policy entry and exit timestamps
preserved exactly.
Policy performance
Overall strategy metrics
| Scope | Trades | Net Profit | Scaled Return | Portfolio-Scaled DD | Weighted Return Sum | PF | Win |
|---|---|---|---|---|---|---|---|
| Loading metrics | |||||||
How Portfolio-Scaled Drawdown Works
Raw trade return is not the same as portfolio return. Each selected action receives a portfolio weight from predicted utility and risk, then its contribution becomes trade return times portfolio weight. Daily weights are normalized and capped by ticker, strategy family, and total gross exposure, so a -5% trade at 5% weight contributes about -0.25% to the portfolio.
Portfolio-Scaled DD is calculated from the daily sum of weighted returns for the selected scope, compounded into an equity curve, then measured against its running high. The older unweighted trade-sequence drawdown is kept only as a diagnostic and is not shown as normal portfolio drawdown.
Candles / entry / exit
Loading chart
Trade examples
Grouped by strategy
Chart data
What the chart shows.
These charts show selected out-of-sample trades generated by the crypto strategy pipeline. Each example comes from the adaptive crypto policy export and uses the exact entry and exit timestamps recorded by the backtest, without recalculating or adjusting them for display.
The price data is pulled from the research database using crypto minute-level OHLCV bars, then resampled into 1-hour candles to make the trade structure easier to inspect. Entry markers show where the policy opened the position, and exit markers show where the recorded trade lifecycle ended. The charts are grouped by strategy family so we can visually compare how different sleeves behave across market regimes.
The policy return comes directly from the adaptive crypto export as
reward_net_ret_chosen. The dashed horizontal levels are displayed
1-hour OHLCV reference closes, so they are visual inspection marks rather than
authoritative execution prices.
Optimization objective
What each strategy type tries to maximize.
The system is built as a two-stage optimization problem. Strategy families first search for indicator conditions that produce favorable forward net returns. A policy model then scores the candidate trades that fire at each decision time and selects the highest-utility candidates subject to portfolio risk constraints.
S* = argmax_S E[R_net(t,h) | S_t = 1]S_t is the active signal condition, h is the holding horizon, and R_net is forward return after transaction costs.
Score(S) =
z(mu_net)
- z(PF)
- z(HR)
- z(Stability)
- z(DD)
- z(Concentration)The rank balances mean net return, profit factor, hit rate, stability, drawdown, and concentration risk.
U_hat(i,t) = f_theta(X_i,t, Z_t)X contains candidate features, signal context, side, exits, and strategy family. Z contains regime context such as volatility, trend, ATR, and relative strength.
U(i,t) =
R_net(i,t)
- alpha * MFE(i,t)
- beta * |MAE(i,t)|
- gamma * StopPenalty(i,t)The target rewards realized net return while penalizing path risk, adverse excursion, and stop behavior.
Portfolio selection
max_w sum_i w_i * U_hat(i,t)
subject to:
sum_i |w_i| <= GrossExposure
sum_{i in ticker_j} |w_i| <= TickerCap_j
sum_{i in strategy_k} |w_i| <= StrategyCap_k
w_i = 0 if U_hat(i,t) < q_tau(U_hat_t)In words, the model selects trades with the highest predicted risk-adjusted utility, while controlling exposure by ticker, strategy family, and total portfolio risk.