Skip to main content

2 posts tagged with "research"

View All Tags

Katana vs Built‑in TPSL – Toward a Fair Objective

· 2 min read
Max Kaido
Max Kaido
Architect

Why “better” needs a clear objective​

When comparing TPSL methods (built‑in heuristic vs Katana combinatorics), plain weighted R:R can reward far TPs and penalize robust SLs. We need an objective that reflects utility under uncertainty, not just geometry.

Observed asymmetries​

  • Built‑in can nudge SL continuously (e.g., to meet R:R≥1.5), while Katana currently samples discrete SL candidates (swing/ATR/heuristic). This can yield higher R:R on built‑in without necessarily overfitting.
  • Allocation grid: 50/50 sometimes wins; Katana limits TP1 to 30–50%. If RR2 uplift is modest, shifting to 30/70 can hurt the weighted result.
  • Precision/rounding: Minor differences move RR.

Example (MYRIAUSDT): built‑in with SL 0.00093309 (50/50) beats Katana using SL 0.00091865 (30/70). The difference stems from SL granularity and allocation grid.

A fair, robust objective​

Replace naive weighted R:R with risk‑aware utility variants:

  • Harmonic mean: penalizes imbalance across TPs.
  • Concave utility: utility = Σ alloc_i · sqrt(RR_i).
  • Capped RR: utility = Σ alloc_i · min(RR_i, RR_cap) to avoid chasing distant TP2.

Penalties (configurable):

  • Risk cost: − α · riskPercent (discourage very wide SLs).
  • Distance cost: − β · max(0, tp2DistancePct − k) (penalize far, low‑probability targets).

Candidate and allocation policy​

  • SL candidates: swing lows, ATR 1.5×, heuristic; optional “RR‑tuned within band” (off by default) to keep parity with built‑in continuous nudge.
  • TP candidates: ATR 1.5×, ATR 2.5×, H1 VWAP; add D1 swing if available.
  • Allocations: TP1 ∈ [30%, 60%] step 5; 2‑ or 3‑level with min 10% on TP3.
  • Constraints: no‑trade ±1%; SL below entry and inside ATR band; ordered TP levels.

Selection and stability​

  • Stability filter: prefer combos that stay top‑quartile across rolling windows.
  • Tie‑breakers: lower riskPercent, then semantic SL (swing > ATR) for explainability.

Implementation plan (incremental)​

  1. Add utility variants to Katana (harmonic, concave, capped).
  2. Optional RR‑tuned SL candidate (disabled by default) and exact rounding parity.
  3. Expand allocation sweep to 30–60% for TP1.
  4. Report columns: builtin_utility, katana_utility_[variant], risk%, flags (rr_tuned_used).
  5. Run across full universe; compare win‑rates per variant and by regime.

Expected outcome​

We trade single‑run optimality for robustness: fewer fragile wins, more consistent, explainable decisions that respect risk. If built‑in still “wins,” it will do so under the same objective—then we adopt that behavior into Katana transparently.