Aktien- und Optionssignale: Backtest-Konfiguration

# setup.yaml


```yaml
strategy_id: sp500_equity_option_analytics
setup_version: v1

universe:
  n_assets: 633
  eligibility_rule: sp500_with_options

decision:
  cadence: weekly_friday_close
  snapshot: friday_16:00_et
  execution_delay: monday_open
  iv_feature_lag: 1_day
  # Per-label decision cadence; `cadence` above applies to every label not named here. The five
  # 5-day labels trade the weekly grid, which is what they forecast. The two 10-day labels
  # resolve after two weeks, so on the weekly grid a new position is opened while the previous
  # one is still inside its own horizon - overlapping the very quantity being measured. They
  # rebalance every other week instead.
  cadence_by_label:
    fwd_ret_10d: biweekly
    fwd_dir_10d: biweekly

# The feature-specification register for 03_financial_features, and every window
# it computes over. It lives here rather than in the notebook for the same reason
# the label name and the holdout boundary do: it is the statement of what this
# case study's feature set is, and a statement only the notebook holds cannot be
# read by a test, by a later stage, or by anyone asking what changed.
#
# ``lookback`` is counted in daily bars back from the decision timestamp and is
# the floor the warmup audit holds each column to. ``lag`` is the delay with
# which the input becomes knowable: every option-derived family carries one
# session because the surface summary is stamped at the close it summarizes and
# is not read until the next session's decision, and every equity-price family
# carries zero because the close is the decision snapshot itself.
features:
  # Surface contract selection, fixed by data/equities/market/sp500/materialize_options.py.
  # Repeated here only so section B can state the observability of what it loads.
  surface:
    dte_buckets: {'7d': [5, 10], '30d': [25, 35], '90d': [80, 110]}
    delta_targets: {atm: 0.50, '25d': 0.25, '10d': 0.10}
  windows:
    iv_zscore: [63, 252]
    iv_percentile: 252
    iv_momentum: [5, 21]
    skew_zscore: 63
    term_zscore: 63
    vrp_zscore: 63
    realized_vol: [20, 63]
    garman_klass: 21
    vol_of_vol: 21
    realized_skew: 21
    momentum: [5, 21, 63, 126, 252]
    skip_recent: 21          # skip-month momentum runs t-252 to t-21
    skip_start: 252
    risk_adjusted: 63        # the momentum horizon its own volatility scales
    # Annualized volatility floor in the risk-adjusted momentum denominator. A share
    # that realized almost nothing over a quarter would otherwise carry an unbounded
    # ratio, and one such row dominates the within-date percentile taken from it.
    risk_adjusted_vol_floor: 0.01
    iv_forward_fill: 5       # sessions a lagged surface value may be carried forward
  # Source column -> the name its within-date percentile is written under. The
  # names predate this register and later stages select by them, so the mapping
  # is explicit rather than a suffix rule.
  ranked:
    iv_30_atm: iv_rank
    skew_rr_30_25d: skew_rank
    ivrv_spread: vrp_rank
    mom_21d: mom_21d_rank
    mom_63d: mom_63d_rank
    rv_20: rv_rank
    iv_mom_21d: iv_mom_rank
    d_iv_30_atm: d_iv_rank
  families:
    - name: cross-sectional rank
      pattern: '*_rank'
      role: signal
      hypothesis: A long-short book acts on relative standing, not on the level of a quantity
      inputs: the eight level and dynamics columns named under features.ranked
      lookback: 252
      lag: 1
      frame: cross section within the decision date
      representation: percentile within the date, in (0, 100)
      failure_mode: a thin cross-section makes the percentile a coarse ordering of few names
    - name: implied volatility level
      pattern: iv_30_atm|iv_7_atm|iv_90_atm|iv_30_put_25d|iv_30_call_25d
      role: signal
      hypothesis: What the option market charges for a name's coming month is priced against what its shares then do
      inputs: daily IV surface summary, delta-selected within fixed DTE buckets
      lookback: 1
      lag: 1
      frame: time series
      representation: annualized implied volatility, in variance points
      failure_mode: a name whose surface is quoted thinly carries a level set by one stale contract
    - name: implied volatility dynamics
      pattern: d_iv_30_atm|iv_mom_*|iv_30_atm_z_*|iv_30_atm_pct_*
      role: signal
      hypothesis: Where implied volatility sits against its own recent history says more than its level
      inputs: at-the-money 30-day implied volatility
      lookback: 252
      lag: 1
      frame: time series
      representation: first difference, multi-session change, trailing z-score and trailing percentile
      failure_mode: the z-score is unbounded as a name's own dispersion approaches zero
    - name: skew and term structure
      pattern: skew_*|term_*|d_skew_*|d_term_*
      role: signal
      hypothesis: The shape of the surface prices asymmetry and horizon that its level cannot express
      inputs: 25-delta put and call IV, and the 7-, 30- and 90-day at-the-money points
      lookback: 63
      lag: 1
      frame: term structure
      representation: differences and ratios between surface points, and their trailing z-scores
      failure_mode: the far segment reads a bucket whose contracts were interpolated rather than quoted
    - name: variance risk premium
      pattern: ivrv_spread|vrp_z_*
      role: signal
      hypothesis: A name whose implied volatility stands above what it goes on to realize is richly priced
      inputs: at-the-money implied volatility and trailing realized volatility
      lookback: 63
      lag: 1
      frame: time series
      representation: spread in volatility points and its trailing z-score
      failure_mode: it compares a forward-looking month against a backward-looking one, which differ around an event
    - name: realized volatility
      pattern: rv_20|rv_63|gk_vol_*|vol_of_vol_*|realized_skew_*
      role: state
      hypothesis: Dispersion and the asymmetry of the realized path set what a ranking can earn
      inputs: split- and dividend-adjusted OHLC bars
      lookback: 63
      lag: 0
      frame: time series
      representation: annualized standard deviation, a range-based estimator, and third moments
      failure_mode: close-to-close misses the overnight gap the range-based estimator is here to catch
    - name: equity momentum
      pattern: mom_*
      role: signal
      hypothesis: Option-derived information has to earn its place beside the price signal it would replace
      inputs: split- and dividend-adjusted close
      lookback: 252
      lag: 0
      frame: time series
      representation: simple return at five horizons, its skip-month form, and its volatility-scaled twin
      failure_mode: reverses over the most recent month, which skip-month momentum drops
    - name: surface quality
      pattern: spread_atm_*|qc_converged_share
      role: state
      hypothesis: A surface point solved from wide or unconverged quotes should not be read like one that was not
      inputs: bid-ask spread and solver convergence flags of the selected contracts
      lookback: 1
      lag: 1
      frame: time series
      representation: relative spread, and the share of selected points that converged
      failure_mode: it measures quotation quality, not liquidity, and the two part company around expiry

# Engine-level execution defaults. Single source of truth — Ch16-19 notebooks
# read these via get_backtest_config(); never declare a local INITIAL_CASH
# or share_type constant. Changing values here invalidates every existing
# backtest_hash for this case study.
#
# Spec-hash inputs (each invalidates every backtest_hash on change):
#   - execution.initial_cash
#   - execution.share_type
#   - execution.allocator_lookback (CS-level fallback for moment allocators)
#   - per-allocator overrides in backtest.sweep.allocators
#
# ``allocator_lookback`` is the bars-of-underlying-price window applied
# uniformly to every moment-based allocator. Daily S&P 500 equity bars
# → 63 ≈ 3 months. ``mvo_ledoit_wolf`` carries an explicit per-allocator
# override below (126 bars ≈ 6 months) so N/K shrinkage degeneracy doesn't
# collapse it to identity-target at top_k=20.
#
# ``initial_cash`` restored to 1_000_000 (2026-05-16) — the 2026-05-15 SSOT
# migration drop to 100k caused near-zero rebalancing here (min num_trades
# = 27 over 4y daily on 500-name × top_k=20: at $5k/name budget × integer
# rounding × the S&P high-priced tail, the engine almost never trades).
# See memory/feedback_2026_05_15_equity_sizing_invalidated.md.
execution:
  initial_cash: 1_000_000        # restores prior validated state
  share_type: integer            # US equities trade in whole shares
  allocator_lookback: 63         # 3 months of daily bars (IV/RP/HRP fallback)

mapping:
  class: long_only_rank_and_rebalance
  position_state_space: long_only
  entry_logic: rank_by_iv_signal
  sizing: equal_weight

costs:
  class: material
  model: percentage          # Loader path: bps regime via per_leg_cost_bps_range midpoint.
  components: [spread, commission, market_impact]
  per_leg_cost_bps_range: [3, 10]
  round_trip_cost_bps: 13    # Midpoint of per_leg_cost_bps_range (6.5 bps × 2 legs).
  # per_share is the commission rate for the exploratory per-share
  # cost-sensitivity regime (read by Ch18 17_costs.py and the run_sweep
  # planner). IBKR Pro Tiered top tier. NOT used in the headline bps
  # regime — bps regime is the production cost model.
  per_share: 0.0035
  note: Trades equities (not options); S&P 500 names are liquid and weekly rebalancing keeps turnover moderate.

backtest:
  rebalance:
    # A rebalance is skipped when the per-asset weight change is below
    # min_weight_change AND the resulting trade notional is below
    # min_trade_value. The benchmark profile disables thresholds so that
    # full-universe equal-weight (1/N per asset) rebalances at all.
    default:
      min_weight_change: 0.005
      min_trade_value: 100.0
    benchmark:
      min_weight_change: 0.0
      min_trade_value: 0.0
  sweep:
    # Iteration controls per stage. ``signal: 0`` means "all predictions";
    # downstream stages take the top-N from the upstream stage's rank-1.
    # Notebooks read these via get_top_n_predictions(case_study, stage).
    top_n_predictions:
      signal: 0                 # all signal predictions (eq-weight baseline)
      allocation: 10            # top-10 model configs by equal-weight baseline Sharpe
      cost_sensitivity: 1       # top-1 of {signal+allocation} per label
      risk_overlay: 1           # top-1 of {signal+allocation} per label
    # Skip MVO/HRP when allocator runtime is the bottleneck (intraday CSes).
    expensive_allocators_skip: false
    # Equal weight is the baseline above, not an allocation-stage method.
    # Score-weighted and inverse-vol take the cheap path.
    # Ch16 signal-stage selection. Long-only equal-weight top-k for every
    # label.
    top_k_grid:
      fwd_ret_5d:           [5, 10, 20]
      fwd_ret_10d:          [5, 10, 20]
      fwd_ret_risk_adj_5d:  [5, 10, 20]
      fwd_dir_5d:           [5, 10, 20]
      fwd_dir_10d:          [5, 10, 20]
    # Ch17 portfolio: reuses top_k_grid above and sweeps over allocators.
    # Moment-based allocators (IV/RP/HRP) use the CS-level
    # ``execution.allocator_lookback`` (63 bars). ``mvo_ledoit_wolf`` gets
    # an explicit 126-bar override (6 months) so N/K ≥ 2.5 at top_k=20 and
    # Ledoit-Wolf shrinkage doesn't collapse to identity-target.
    # No max_weight cap — see memory/feedback_max_weight_caps_intentionally_absent.md
    # (the prior 0.40 cap was already the loosened-from-0.20 workaround;
    # restoring it would push moment allocators back toward equal-weight).
    allocators:
      - {name: score_weighted,  method: score_weighted}
      - {name: inverse_vol,     method: inverse_vol}
      - {name: risk_parity,     method: risk_parity}
      - {name: mvo_ledoit_wolf, method: mvo_ledoit_wolf, lookback: 126}
      - {name: hrp,             method: hrp}
      # Sizes each position by the width of its conformal prediction interval rather than by a
      # moment of returns, so it is the one allocator here that reads the model's own
      # uncertainty. It was absent while etfs, cme_futures and fx_pairs declared it, which made
      # this case study's sweep narrower than theirs for no stated reason. 13_model_analysis
      # measures the coverage those widths come from, so a run of this allocator is also the
      # test of whether that calibration is good enough to size with.
      - {name: conformal_weighted, method: conformal_weighted}
    # Ch18 cost sensitivity (bps regime — headline). A companion per-share
    # regime is run from Ch18 cost notebooks for regime comparison; for
    # sp500_eoa it is exploratory only because flat-default half-spread on
    # split-adjusted prices conflates split adjustment with realized
    # friction.
    cost_grid_bps: [0, 1, 2, 3, 5, 7, 10, 15, 20, 30, 50]
    # Companion per-share half-spread grid (USD per share). Swept alongside
    # cost_grid_bps by the planner. Values: 0¢, 0.5¢, 1¢, 2.5¢, 5¢, 10¢.
    cost_grid_half_spread_usd: [0.0, 0.005, 0.01, 0.025, 0.05, 0.10]
    # Ch19 risk overlays.
    risk_controls:
      position:
        - {name: stop_loss_3pct,  type: stop_loss,     threshold: 0.03}
        - {name: stop_loss_5pct,  type: stop_loss,     threshold: 0.05}
        - {name: stop_loss_10pct, type: stop_loss,     threshold: 0.10}
        - {name: stop_loss_15pct, type: stop_loss,     threshold: 0.15}
        - {name: trailing_1pct,   type: trailing_stop, threshold: 0.01}
        - {name: trailing_2pct,   type: trailing_stop, threshold: 0.02}
        - {name: trailing_3pct,   type: trailing_stop, threshold: 0.03}
        - {name: trailing_5pct,   type: trailing_stop, threshold: 0.05}
        - {name: trailing_10pct,  type: trailing_stop, threshold: 0.10}
        - {name: trailing_15pct,  type: trailing_stop, threshold: 0.15}
        - {name: trailing_20pct,  type: trailing_stop, threshold: 0.20}
        - {name: time_exit_10,    type: time_exit,     bars: 10}
        - {name: time_exit_20,    type: time_exit,     bars: 20}
        - {name: time_exit_40,    type: time_exit,     bars: 40}

evaluation:
  n_splits: 2
  train_size: 2Y
  val_size: 1Y
  holdout_start: '2021-01-01'
  holdout_end: '2021-12-31'
  calendar: NYSE
  periods_per_year: 252  # NYSE 5d/wk

labels:
  primary: fwd_ret_5d
  buffer: 10D
  # Outcome horizons seal validation before holdout. They are separate from
  # the deliberately conservative primary train-to-validation buffer above.
  horizons:
    fwd_ret_5d: 5D
    fwd_ret_10d: 10D
    fwd_ret_risk_adj_5d: 5D
    fwd_dir_5d: 5D
    fwd_dir_10d: 10D
  variants:
    - fwd_ret_10d
    - fwd_ret_risk_adj_5d
    - fwd_dir_5d
    - fwd_dir_10d
  variant_buffers:
    fwd_ret_10d: 10D
    fwd_ret_risk_adj_5d: 5D
    fwd_dir_5d: 5D
    fwd_dir_10d: 10D
  # Vectorized-backtest thinning step per label: number of schedule slots
  # to advance per trade so holding periods don't overlap.
  # The step is ceil(horizon / cadence), and `cadence` is now per label, so it must be read
  # against decision.cadence_by_label rather than against the weekly default. The two 10-day
  # labels moved to the biweekly grid, which already spaces their decisions 10 sessions apart;
  # leaving their step at 2 would thin an already-thinned schedule and trade them every four
  # weeks under a spec that says biweekly.
  rebalance_step:
    fwd_ret_5d: 1            # ceil(5 / 5) on the weekly_friday_close schedule
    fwd_ret_10d: 1           # ceil(10 / 10) on the biweekly schedule
    fwd_ret_risk_adj_5d: 1   # ceil(5 / 5)
    fwd_dir_5d: 1            # ceil(5 / 5)
    fwd_dir_10d: 1           # ceil(10 / 10) on the biweekly schedule
  # Continuous return that each classification label is derived from.
  classification_eval_label:
    fwd_dir_5d: fwd_ret_5d
    fwd_dir_10d: fwd_ret_10d

modeling:
  gbm:
    libraries: [lightgbm]
    preset: default
    device: cpu
    # LightGBM's own CPU default. 63 is the GPU default and was carried over with the
    # device when these runs moved off the GPU, so every CPU fit was quantizing the design
    # matrix into a quarter of the bins the library would have used. Coarser bins are
    # faster and lose split points; the reader running this on a CPU gets what the
    # documentation describes.
    max_bin: 255
  latent_factors:
    persistent_entities: true
    # The device the family is fitted on. Declared rather than left out: it sits inside
    # the hashed computation (case_studies/utils/latent_factors/adapter.py, computation
    # .runtime and .numerical_runtime), and with no key here case_study.py falls back to
    # preferred_latent_device(), which returns cuda or cpu depending on what the machine
    # running it happens to have. That makes the training identity a property of the host.
    # The three neural members are fitted here; 11a_pca and 11b_ipca override it to cpu,
    # because run_pca_fold and run_ipca_fold are numpy and scipy and take no device at
    # all, so recording cuda for them would describe a computation that did not happen.
    device: cuda
    model_kwargs:
      ipca:
        max_iter: 10000
        factor_ridge: 0.01
        gamma_ridge: 0.01
      sdf:
        checkpoint_epochs: [256, 512, 768, 1024]  # conditional-relative; publishes global 256..1280
        beta_checkpoint_epochs: [256]
        beta_default_checkpoint: 256
      sae:
        # Rows per gradient step. Declared because the alternative is not a smaller batch but
        # no batching at all: SAEConfig.batch_size defaults to None and the library reads that
        # as one batch holding the whole training window, about 250,000 rows here, which does
        # not fit a 24 GB card. Its sibling run_cae_fold has carried this same value as a
        # runner default all along; only the SAE runner never passed one.
        batch_size: 10000

# What `04_model_based_features` decides, declared here for the same reason the feature
# windows are: an estimation window is part of a fitted feature's definition, so it belongs
# where the definition lives rather than inside the notebook that runs it. Every count is in
# NYSE sessions.
#
# A fitted feature is bounded by the schedule below and not by a cross-validation fold. The
# parameters behind a value are estimated from sessions strictly before it, refreshed on the
# cadence given, and the same value comes out whichever fold later selects the row - so the
# artifact carries no fold column.
#
# The panel runs 2017-01-03 to 2021-12-31, 1,259 sessions over 624 securities carrying an
# option surface. Fold 0's training window opens 2018-01-04, so only about 250 sessions of
# history precede the first fold, and that run-up is what a burn-in has to be paid out of.
model_based:
  gjr_garch:
    # Sessions of a security's own adjusted returns before its volatility model is fitted.
    # One year, and it is bounded above by the panel rather than chosen freely: 252 is very
    # nearly the 250 sessions that precede fold 0, so it comes out of the run-up instead of
    # out of training data. Measured against the alternative on the 624-security roster, a
    # 504 burn-in takes emitted coverage from 76.5% of the panel to 54.7%, takes the
    # securities that emit nothing at all from 57 to 103, and eats 159,463 of fold 0's
    # training sessions across 592 securities rather than 22,130 across 121.
    #
    # It is not free. A GJR-GARCH fitted on 252 observations returns a degenerate parameter
    # vector - alpha + gamma < 0, so a larger down-shock lowers next session's variance -
    # for 19.0% of securities on the first block, against 1.8% of fits averaged over the
    # whole walk as the expanding window grows. Section C reports both.
    burnin: 252
    # How often the parameters are re-estimated. A month. Each estimate is a
    # quasi-maximum-likelihood fit over the whole expanding history, and the leverage
    # parameter it is estimated for is not a fast-moving quantity.
    refit_every: 21
    # The specification itself, for the same reason the schedule is here: what was fitted is
    # part of what the feature means, and a reader comparing this chapter's model-based
    # features against another's needs to see where they differ without reading two notebooks.
    #
    # `o: 1` is this case study's declared deviation from the shared GARCH(1,1)-Normal default,
    # and it is what makes this a GJR rather than a plain GARCH. The justification is a property
    # of the data, not a preference: a negative return raises next session's variance by more
    # than a positive return of the same size, and on single-name equity underlying an option
    # book that asymmetry is what the option prices are quoted around. Section C of the notebook
    # carries the measurement. `sp500_options` declares the same deviation for the same reason.
    mean: Constant
    vol: GARCH
    p: 1
    o: 1
    q: 1
    dist: Normal

causal:
  treatment: ivrv_spread
  # Bars the treatment's own construction window spans, which is what the placebo block has
  # to cover: permuting ivrv_spread in blocks shorter than this destroys the serial
  # dependence the refutation exists to preserve, and the resulting p-value reads like a
  # refutation without being one. Declared here rather than inferred, because guessing which
  # element of a window list a column was built from puts a wrong number behind a right-looking
  # one. Derived from the construction, not chosen:
  #
  # `iv_30_atm - rv_{realized_vol[0]}` in 03_financial_features, and features.windows
  # declares realized_vol as [20, 63]. The implied leg is a quote and rolls nothing; the
  # realized leg is what makes the spread autocorrelated, over its own 20 sessions.
  treatment_window: 20
  confounders: [rv_20, mom_21d, skew_rr_30_25d]
  method: walk_forward_dml

```