Konfiguration eines Long-Short-Backtests für US-Aktienpanels

# setup.yaml


```yaml
strategy_id: us_equities_panel
setup_version: v1

decision:
  cadence: daily_close
  snapshot: close
  execution_delay: next_bar_open

universe:
  n_assets: 3199

# 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 (e.g. lookback,
#     vol_window, max_weight)
#
# ``allocator_lookback`` is the bars-of-underlying-price window applied
# uniformly to every moment-based allocator. Daily US 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=50.
#
# ``initial_cash`` restored to 1_000_000 (2026-05-16) — the 2026-05-15 SSOT
# migration drop to 100k caused catastrophic degenerate output here (top_k=50
# leg suspect; integer-share rounding × high-priced names). 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_short_decile_rebalance
  position_state_space: long_short
  entry_logic: decile_sort_long_top_short_bottom
  sizing: equal_weight_within_decile

costs:
  class: material
  model: percentage          # Loader path: bps regime via per_leg_cost_bps_range midpoint.
  components: [spread, commission, market_impact, borrow_cost]
  per_leg_cost_bps_range: [5, 20]
  # per_share is the commission rate for the exploratory per-share
  # cost-sensitivity regime (read by Ch18 19_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
  # Documentation-only: loader reads top-level per_leg_cost_bps_range. The
  # era split is qualitative — see feasibility §B.4 for why flat bps suffices
  # given the validation window starts 2000-01-12 (pre-decimal era is ~6.5%
  # of the window).
  era_dependent:
    pre_decimalization:
      period: before 2001-01-29
      per_leg_cost_bps_range: [15, 30]
      note: Tick size 1/16 ($0.0625); wider spreads, higher commissions.
    post_decimalization:
      period: after 2001-01-29
      per_leg_cost_bps_range: [5, 15]
      note: Penny tick regime; electronic trading, lower spreads.
  borrow_cost_note: Long-short requires borrow for the short leg (~50 bps/yr).

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
    # All others (equal_weight, score_weighted, inverse_vol) take the cheap path.
    # Ch16 signal-stage selection. Long-short equal-weight top-k on three
    # label horizons (1d primary, 5d/21d variants). The wider top-k grid
    # reflects the larger universe (~3000 names vs ETFs' 100).
    top_k_grid:
      fwd_ret_1d:  [20, 50]
      fwd_ret_5d:  [20, 50]
      fwd_ret_21d: [20, 50]
    # 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=50 and
    # Ledoit-Wolf shrinkage doesn't collapse to identity-target.
    # SKIP_EXPENSIVE_ALLOC filters mvo_ledoit_wolf and hrp at notebook
    # level when requested.
    # No max_weight cap — see memory/feedback_max_weight_caps_intentionally_absent.md
    # (capping near top_k forces equal-weight and defeats the comparison).
    allocators:
      - {name: equal_weight,    method: equal_weight}
      - {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}
      - {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
    # us_equities_panel it is exploratory only because the wide price
    # distribution + adjusted-price confounder make a flat-default
    # half-spread an artifact rather than realistic 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: 16
  train_size: 10Y
  val_size: 1Y
  holdout_start: '2016-01-01'
  holdout_end: '2018-03-31'
  calendar: NYSE
  periods_per_year: 252  # NYSE 5d/wk

labels:
  primary: fwd_ret_1d
  buffer: 1D
  variants:
    - fwd_ret_5d
    - fwd_ret_21d
  variant_buffers:
    fwd_ret_5d: 5D
    fwd_ret_21d: 21D
  # Vectorized-backtest thinning step per label: number of schedule slots
  # to advance per trade so holding periods don't overlap.
  rebalance_step:
    fwd_ret_1d: 1
    fwd_ret_5d: 5
    fwd_ret_21d: 21

model_based:
  regime:
    # Calm months and falling ones. Two states is the coarsest split that separates the
    # calm, mildly positive months from the falling, turbulent ones, which is the
    # distinction the momentum-crash literature works in.
    n_clusters: 2
    # Sessions per clustered window, and how many sessions two consecutive windows share.
    # A month of sessions per window, overlapping by a week, so a shift in the return
    # distribution is seen by more than one window before it is clustered.
    window: 21
    overlap: 5
    # Sessions of market history spent before the first clustering is fitted. Three years.
    # At the 16-session step the window and overlap imply, that is 46 windows to cluster,
    # which is the fewest that separates two centroids from the noise in a single window.
    # The clustered series is one market-level series rather than a panel, so this is paid
    # once for the whole notebook. It matches etfs, whose regime model reads a daily
    # market-level series of the same shape.
    burnin: 756
    # How often the centroids are re-estimated. A quarter. Regime parameters are the
    # slowest-moving thing this notebook fits and each estimate is n_init k-means searches
    # over the whole history so far.
    refit_every: 63
  garch:
    # Sessions of a stock's own returns before its variance model is fitted. Two years.
    # Below that a maximum-likelihood fit of three parameters returns estimates whose
    # standard errors swamp them, and the persistence term - the one the feature is most
    # sensitive to - is estimated from too few volatility cycles.
    burnin: 504
    # How often the variance model is re-estimated. A quarter, which on this panel is
    # 3,160 series - the rest carry less than the burn-in above and take the market-level
    # volatility - re-estimated 58 times each on average, 110 times for a series that
    # spans the whole sample, and about 183,000 times in all. Section 4b measures what the
    # cadence buys: the fitted persistence of a US equity moves slowly, so a faster cadence
    # would multiply the fits without moving the emitted volatility.
    refit_every: 63
    # The specification the maximum-likelihood fit estimates, declared here rather than
    # written into the notebook's call, so what "the GARCH feature" names on this panel is
    # readable without opening the code. These five are the argument names `arch_model` takes
    # and 04 passes them through unchanged.
    #
    # This is the plain GARCH(1,1)-Normal baseline: a constant mean, a symmetric variance
    # recursion of order (1,1), and Gaussian innovations. No key here differs from that
    # baseline, so this case study declares no deviation.
    #
    # `o: 0` is the symmetric model. Setting it to 1 adds the leverage term - down days
    # raising next-day variance more than up days of the same size - which is what
    # sp500_options and sp500_equity_option_analytics fit on their index series. It is not
    # fitted here because a broad cross-section of single stocks is where the term is least
    # reliably identified: many names carry too few large down moves in a two-year estimation
    # window to separate gamma from alpha, and Section 4b's persistence spread is already the
    # widest thing the schedule has to hold steady.
    #
    # `dist: Normal` is the likelihood, not the filter. It changes which coefficients the
    # optimizer returns and changes no step of the variance recursion those coefficients are
    # then run through.
    mean: Constant
    vol: GARCH
    p: 1
    o: 0
    q: 1
    dist: Normal

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
    # Down-tune the SDF schedule on this large-N CS so training time stays
    # in the overnight budget. Paper defaults (256/64/1024) on broad US
    # equities at daily frequency would take many hours per fold.
    model_kwargs:
      ipca:
        max_iter: 10000
        factor_ridge: 0.01
        gamma_ridge: 0.01
      sdf:
        n_epochs_unc: 128
        n_epochs_moment: 32
        n_epochs_cond: 512
        burn_in_epochs: 32
        checkpoint_epochs: [128, 256, 384, 512]  # conditional-relative; publishes global 128..640
        beta_checkpoint_epochs: [256]
        beta_default_checkpoint: 256

causal:
  treatment: past_ret_12m_skip
  # Bars the treatment's own construction window spans, which is what the placebo block has
  # to cover: permuting past_ret_12m_skip 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:
  #
  # `close.shift(MOMENTUM_SKIP) / close.shift(MOMENTUM_LOOKBACK) - 1` in 02_labels, where
  # those are 21 and 252 trading sessions. 03_financial_features recomputes it as
  # `ret_12m_skip`; the oldest price either reads is 252 sessions back.
  treatment_window: 252
  confounders: [vol_21d, illiq_rank, volume_ratio]
  method: walk_forward_dml

# Declared here and cited in prose, not loaded by any notebook. 01_feasibility_analysis
# section C.2 names the four thresholds and says where each one is measured; no code reads
# this block, so nothing gates on it.
kill_conditions:
  ic_floor: 0.01
  ic_floor_note: Cross-sectional IC below 0.01 across all features.
  edge_to_cost_floor: 1.2
  edge_to_cost_note: Net Sharpe / cost ratio below 1.2x.
  micro_cap_concentration: 0.5
  micro_cap_note: Alpha concentrated >50% in bottom ADV quintile (untradeable).
  net_sharpe_floor: 0.3
  net_sharpe_note: Net Sharpe after borrow costs below 0.3.

```