3x S&P ensemble frontier — walk-forward confirmation

kind

user_hypothesis

issue

1312

parent issue

1310

related issue

1313

primary window

1999-03-05 to 2026-07-24

objective

preregister expanding-window selection and test whether the feasible CAGR/drawdown frontier survives without full-history parameter choice

• Full-window CAGR must be north of 30%, max drawdown must be no worse than -25%, and the only risky holding is levered 3x S&P (UPRO/cash).

• Analyze the parent frontier's large drawdowns and identify indicators that would have avoided those episodes.

• Aggressive in-sample and hindsight overfitting is explicitly in scope, but the instruction does not turn future-return leakage into a deployable rule.

• Evaluate causal full-window fits separately from memorized-event and forward-return oracle ceilings.

• Use trend, shock, volatility, drawdown, breadth, macro, vote, veto, and calendar/oracle exposure maps to span conceptually diverse defenses.

• Realized UPRO history before 2009, exact point-in-time constituent breadth, and prospective performance of a hindsight-fit rule are unavailable.

• Full-window CAGR must be north of 30%, max drawdown must be no worse than -25%, and the only risky holding is levered 3x S&P (UPRO/cash).

• Analyze the parent frontier's large drawdowns and identify indicators that would have avoided those episodes.

• Aggressive in-sample and hindsight overfitting is explicitly in scope, but the instruction does not turn future-return leakage into a deployable rule.

• Evaluate causal full-window fits separately from memorized-event and forward-return oracle ceilings.

• Use trend, shock, volatility, drawdown, breadth, macro, vote, veto, and calendar/oracle exposure maps to span conceptually diverse defenses.

• Does the issue-1310 causal UPRO/cash feasible frontier survive an expanding annual January walk-forward that freezes parameters from training prefixes only?

inputs

• SPY

• UPRO

• TB3MS

• UNRATE

• __SYNTH_SPXBREADTH50__

market sqlite sha256

1537834d27b92bbf1872d98708186870e59d24f5de7a95728a3e08c959d7cff3

synthetic method

backtest.sim_history.synthetic_levered_daily/testfol_daily-v1

synthetic parameters

leverage

3.0

swap exposure

1.1

financing spread percent

0.4

expense ratio percent

1.0

annualizer

252

causal execution

close-derived decision applies one trading bar later

hindsight execution

memorized dates or same-session/forward return; explicitly non-causal and non-deployable

public data policy

publish derived metrics and normalized series, not bulk provider rows or subscriber workbook data

trial count

1873

causal trial count

1816

hindsight trial count

57

acceptance hit count

16

causal acceptance hit count

0

winner variant id

l3x.hindsight.036f423a11de

walk forward issue

1312

walk forward protocol

cost bps

5.0

execution lag bars

1

first cutoff

2010-01-01

fold cadence

annual_january

primary verdict

stitched OOS path max_drawdown >= -0.25

protocol id

expanding-annual-jan-local-c5/v1

schedule

expanding_window

training objective

max CAGR subject to training max_drawdown >= floor

full sample parameter choice

False

variant id

l3x.wf.primary_local_stitched

family id

h1-local-c5-expanding-selection

parameters

protocol id

expanding-annual-jan-local-c5/v1

unique selected

• l3x.c5_vol_scaled_ensemble.547a9b4d36f5

• l3x.c5_vol_scaled_ensemble.e5003d8780f5

created by

walk_forward_confirmation

reason tested

Primary preregistered local expanding-window path

status

exploratory_causal

metrics

cagr

0.12231073495697942

max drawdown

-0.2545491086115719

acceptance hit

False

tags

• interesting_negative_result

• parameter_sensitive

• needs_more_research

• name

  trend_sma_days

  type

  integer

  default

  150

  min

  5

  max

  252

• name

  momentum_days

  type

  integer

  default

  189

  min

  5

  max

  252

• name

  fast_vol_threshold

  type

  float

  default

  0.25

  min

  0.1

  max

  0.3

• name

  breadth_threshold

  type

  float

  default

  0.5

  min

  0.3

  max

  0.7

• name

  oracle_return_cut

  type

  float

  default

  0.0

  min

  -0.01

  max

  0.005

• name

  execution_lag_bars

  type

  integer

  default

  1

  min

  0

  max

  2

• Should confirmation privilege the stitched path or per-fold clearance?

• Would a longer first training window or biennial cadence remove the 0.45pp primary breach without retuning?

• Should the cleared safety-margin rule be forked into a separate indicator campaign after owner review?

• iteration number

  1

  issue

  1310

  objective

  maximize causal exploratory CAGR under the -25% drawdown and UPRO-only constraints

• iteration number

  2

  issue

  1313

  objective

  find >30% full-window hindsight frontier and attribute the parent baseline's large drawdowns

• iteration number

  3

  issue

  1312

  objective

  walk-forward confirmation without full-sample parameter choice

• Expanding annual January selection freezes a causal UPRO/cash rule from each training prefix and applies it one bar later on the next OOS year.

• Primary menu is the 33-variant local neighborhood around the issue-1310 c5 winner; focused grid and -22% training buffer are sensitivities.

• 17 January cutoffs from 2010-01-01; local 33 + focused 161 + safety-margin local candidates.

• Selections never see OOS rows; issue-1313 oracles excluded.

• Safety-margin and focused-grid stitched paths clear -25% (-21.53%, -24.26%).

• Primary annual OOS segments all clear individually; OOS CAGR exceeds the full-history parent winner.

• Primary local stitched path max drawdown -25.45% fails the preregistered floor.

• Full-history winner parameters are not stably reselected on every prefix.

Mixed confirmation: the feasible frontier does not unambiguously survive the primary local protocol, but buffered/broader menus still clear -25% out of sample. Keep exploratory; do not promote.