Chat
Show my 10 worst drawdown days over the last 6 months and what drove each.
code · edited
import pandas as pd from arcana import sdk pf = sdk.portfolio("Global Alpha Fund") # 1 — Daily portfolio returns, trailing 6 months returns = pf.returns(start="2025-12-12", end="2026-06-12", freq="D") # 2 — The 10 worst drawdown days in the window worst = returns.sort_values("return_bps").head(10) # 3 — Position-level contributions on those dates contrib = pf.contributions(dates=worst.index) # 4 — Reconcile: contributions must sum back to the portfolio return. # Tolerance: max(5 bps, 2% of |portfolio return|) — hybrid rule check = contrib.groupby("date")["contribution_bps"].sum().to_frame("sum_bps") check["portfolio_bps"] = worst["return_bps"] check["residual_bps"] = (check["sum_bps"] - check["portfolio_bps"]).abs() check["tolerance_bps"] = pd.concat([ pd.Series(5.0, index=check.index), 0.02 * check["portfolio_bps"].abs(), ], axis=1).max(axis=1) check["reconciled"] = check["residual_bps"] <= check["tolerance_bps"] # 5 — Top 3 contributors per drawdown date top3 = (contrib.sort_values(["date", "contribution_bps"]) .groupby("date").head(3)) print(worst.join(check[["residual_bps", "reconciled"]])) print(top3)
output
No native widget answers this directly. Worst-day identification lives in performance; per-date position attribution lives in contribution analysis — no single surface joins them and reconciles the two. Code Mode ran the join deterministically rather than have the model estimate it. A native answer wasn't available for this question.
Validation for Run 1 · original
These checks reflect Run 1. You've edited the code — rerun to revalidate against your changes.
Data
Daily portfolio returns — Global Alpha Fund · 12 Dec 2025 → 12 Jun 2026
Position-level daily contributions · 10 dates · Arcana performance store (as of 12 Jun 2026, 06:00 ET)
Position-level daily contributions · 10 dates · Arcana performance store (as of 12 Jun 2026, 06:00 ET)
Universe
Positions held on each drawdown date (87–94 names)
Method
Identified the 10 lowest daily portfolio returns in the trailing 6 months, pulled position-level contributions for each date, and reconciled contributions against the portfolio return.
Validation checks
Residual
Changes from Run 1
Reconciled when |Σ position contributions − portfolio return| ≤ max(5 bps, 2% × |portfolio return|). Hybrid tolerance applied internally; constants owned by analytics. Residuals surface above; the formula stays in this disclosure (PRD §8).
What changed from Run 1
You widened the reconciliation tolerance from 5 bps to 10 bps. The 2 dates previously outside tolerance — 07 Apr and 12 Mar — now reconcile, so the caveat is cleared and all 4 checks pass. Worst-day ranking and top contributors are unchanged.
Three episodes account for 8 of your 10 worst days: the early-March semis selloff (4 days, NVDA the largest detractor on each), the mid-April financials drawdown (3 days, JPM-led), and the mid-February mega-cap pullback. NVDA appears among the top contributors on 5 of the 10 dates — your single largest drawdown driver this half.
| Date | Portfolio return | Top contributors (bps) |
|---|---|---|
| 04 Mar 2026 | −212 bps | NVDA −58 · MSFT −41 · AVGO −33 |
| 05 Mar 2026 | −187 bps | NVDA −49 · TSM −38 · AMD −27 |
| 17 Apr 2026 | −166 bps | JPM −44 · GS −31 · MS −24 |
| 13 Feb 2026 | −158 bps | AAPL −39 · AMZN −33 · META −29 |
| 03 Mar 2026 | −141 bps | NVDA −36 · MSFT −30 · GOOGL −22 |
| 16 Apr 2026 | −129 bps | JPM −33 · BAC −26 · C −19 |
| 12 Feb 2026 | −117 bps | AMZN −31 · META −25 · NFLX −18 |
| 07 Apr 2026 | −104 bps | XOM −28 · CVX −22 · COP −17 |
| 21 Jan 2026 | −98 bps | LLY −26 · UNH −21 · JNJ −15 |
| 12 Mar 2026 | −91 bps | NVDA −24 · AVGO −19 · QCOM −14 |
Caveat: Position-level contributions did not fully reconcile with portfolio-level returns for 2 of the 10 drawdown dates, so those rows should be treated as directional. Affected dates are flagged in the table (07 Apr · 12 Mar).
I couldn't complete this computation. Portfolio returns came back fine and the 10 worst days were identified, but the position-contribution request returned no data for 04–05 Mar 2026 even after retries — so the drawdowns can't be attributed reliably. No partial numbers; you can retry now, or rephrase to run without attribution.