株式のイントラデイ・マイクロストラクチャー調査設計

# setup.yaml


```yaml
strategy_id: nasdaq100_microstructure
setup_version: v1

universe:
  # Point-in-time NASDAQ-100 membership over the sample, restricted to the names
  # that contribute at least one session before evaluation.holdout_start. The
  # published AlgoSeek archive carries 123 symbols; eight of them first quote
  # after 2021-07-01, so nothing is ever fit on them, the pre-holdout liquidity
  # profile has no spread for them, and 05_evaluation drops them before it
  # screens anything. 01_feasibility_analysis asserts this list against the
  # archive in both directions. Names that leave the index end where they left:
  # AAL and WLTW stop at 2020-05-01, which is the membership, not a gap.
  symbols:
    - AAL
    - AAPL
    - ADBE
    - ADI
    - ADP
    - ADSK
    - AEP
    - ALGN
    - ALXN
    - AMAT
    - AMD
    - AMGN
    - AMZN
    - ANSS
    - ASML
    - ATVI
    - AVGO
    - BIDU
    - BIIB
    - BKNG
    - BMRN
    - CDNS
    - CDW
    - CERN
    - CHKP
    - CHTR
    - CMCSA
    - COST
    - CPRT
    - CSCO
    - CSGP
    - CSX
    - CTAS
    - CTSH
    - CTXS
    - DLTR
    - DOCU
    - DXCM
    - EA
    - EBAY
    - EXC
    - EXPE
    - FAST
    - FB
    - FISV
    - FOX
    - FOXA
    - GILD
    - GOOG
    - GOOGL
    - IDXX
    - ILMN
    - INCY
    - INTC
    - INTU
    - ISRG
    - JD
    - KDP
    - KHC
    - KLAC
    - LBTYA
    - LBTYK
    - LRCX
    - LULU
    - MAR
    - MCHP
    - MDLZ
    - MELI
    - MNST
    - MRNA
    - MRVL
    - MSFT
    - MTCH
    - MU
    - MXIM
    - NFLX
    - NTAP
    - NTES
    - NVDA
    - NXPI
    - OKTA
    - ORLY
    - PAYX
    - PCAR
    - PDD
    - PEP
    - PTON
    - PYPL
    - QCOM
    - REGN
    - ROST
    - SBUX
    - SGEN
    - SIRI
    - SNPS
    - SPLK
    - SWKS
    - TCOM
    - TEAM
    - TMUS
    - TSLA
    - TTWO
    - TXN
    - UAL
    - ULTA
    - VRSK
    - VRSN
    - VRTX
    - WBA
    - WDAY
    - WDC
    - WLTW
    - XEL
    - XLNX
    - ZM
  n_assets: 115
  eligibility_rule: nasdaq100_membership_with_pre_holdout_session
  # Cost-feasible universe: the cheapest-to-trade names by round-trip cost,
  # frozen per split. Provenance: _build_cost_feasible_universe.py
  # (round-trip proxy 2*(per_share/mean_price)*1e4 + 2*median_half_spread_bps).
  # The validation list is profiled on quote bars strictly before the
  # validation window (2020-01-01 -> 2020-06-30); the holdout list from the
  # pre-holdout liquidity profile (< 2021-07-01). No look-ahead. Selected via
  # strategy.signal.universe_filter='cost_feasible'; the split-specific list is
  # chosen at backtest time by resolving the prediction set's split (NOT a
  # spec-hash input — like the 'liquid' filter, the hash carries only the
  # filter name). The full 115 universe collapses out of sample (featured
  # config holdout -0.42) while the cost-feasible universe is positive; the
  # screen is load-bearing. See 17_costs.py for the full-vs-screened contrast.
  cost_feasible:
    validation:
      - AAPL
      - MSFT
      - FB
      - SBUX
      - PEP
      - ATVI
      - INTC
      - PYPL
      - MDLZ
      - AMD
      - QCOM
      - TXN
      - GILD
      - MU
      - NVDA
      - AMAT
      - CSCO
      - TMUS
      - WBA
      - COST
      - JD
      - CSX
      - XEL
      - CMCSA
      - EXC
      - EBAY
      - MNST
      - FISV
      - FAST
      - EA
      - CTSH
      - NFLX
      - XLNX
      - AMZN
      - ADBE
      - PAYX
      - CDNS
      - AMGN
      - MXIM
      - BIDU
      - PCAR
      - CERN
      - UAL
      - ADP
      - KHC
      - FOXA
      - WDC
      - TSLA
      - FOX
      - CTXS
    holdout:
      - AAPL
      - MSFT
      - FB
      - SBUX
      - PEP
      - GILD
      - AMD
      - QCOM
      - INTC
      - MDLZ
      - ATVI
      - AMAT
      - MU
      - PYPL
      - AEP
      - TXN
      - JD
      - CSX
      - MRVL
      - COST
      - TMUS
      - CMCSA
      - EBAY
      - CSCO
      - NVDA
      - XEL
      - FISV
      - WBA
      - EXC
      - FAST
      - TSLA
      - CTSH
      - MXIM
      - ADI
      - AMZN
      - NFLX
      - AMGN
      - EA
      - PAYX
      - CERN
      - ADBE
      - ADP
      - KHC
      - CDNS
      - XLNX
      - MNST
      - KDP
      - BIDU
      - DLTR
      - UAL

decision:
  # How often the strategy is allowed to act, NOT the size of a bar in the panel.
  # The observations are one minute apart; 03_financial_features keeps every
  # fifteenth of them to build the decision schedule. Nothing here declares the
  # observation grid, so anything that needs it - an embargo, a permutation block,
  # any count of periods spanning a duration - measures it from the data instead
  # of dividing by this number. Reading this field as the bar size yields an
  # answer fifteen times too small and nothing in the result shows it.
  bar_frequency: 15_minute
  # A model decides on its own horizon: a five-minute model every five minutes, a
  # sixty-minute model every hour. `bar_frequency` above is the default for a label that
  # declares nothing here, and stays the primary label's cadence.
  #
  # This is the decision schedule only. The engine's price feed is one minute
  # (`config/backtest/base.yaml::calendar.data_frequency`), so a position entered on one
  # decision is watched every minute until the next - which is what a stop or a
  # take-profit is measured on, and what a fifteen-minute feed could not express.
  cadence_by_label:
    fwd_ret_5m: 5_minute
    fwd_ret_15m: 15_minute
    fwd_dir_15m: 15_minute
    fwd_ret_60m: 60_minute
  decision_snapshot: bar_close
  execution_delay: 1_bar
  execution_price_assumption: next_bar_open_or_vwap

# The feature specification read by 03_financial_features. Windows are counted
# in minute bars and every one of them is session-bounded: a window never spans
# an overnight gap, so each session restarts its own warmup.
features:
  # Feature matrices are stored and fitted in single precision. 16,098,877 rows x 88 features
  # in float64 is a 10.9 GB modeling dataset and a 58.9 GB peak through one gradient-boosting
  # fold; in float32 those are 5.6 GB and 38.0 GB. LightGBM bins to uint8 regardless, and the
  # linear families fit standardised columns where the extra mantissa carries nothing this
  # case study can measure. Declared here rather than globally: the smaller case studies fit
  # comfortably in double precision, and narrowing them would move their numbers for nothing.
  storage_dtype: float32
  windows:
    fast: 5              # one fast aggregate
    decision: 15         # one decision bar at the declared bar_frequency
    slow: 30             # realized-volatility horizon and the Amihud average
    hour: 60             # Kyle's lambda regression and the slow regime average
    ewma_half_life: 30   # half-life of the EWMA volatility, in bars
    stale_cap: 5         # consecutive stale-quote bars tolerated before nulling
    edge_block: 30       # width of the open and close blocks the flags mark
  # The quantity the thesis puts forward: the order-flow imbalance measured over
  # one decision bar. Its one-bar form is this case study's causal treatment (see
  # `causal.treatment` below). Read by 03_financial_features, which draws it in
  # F2, F3 and F6.
  carrier: signed_vol_share_15m
  # One row per family. `pattern` claims a family's columns in both
  # representations - the level and its cross-sectional z-score twin - because
  # the two carry the same hypothesis on two scales. `lookback` and `lag` are
  # minute bars, and they are what the warmup audit and the timing figure read.
  families:
    - name: quote_liquidity
      pattern: rel_spread_*|depth_imb*|quote_rate*
      role: state
      hypothesis: >-
        The cost and the depth of the book describe the environment a signal is read
        in: an imbalance is worth acting on only where the round trip is cheap enough
        to leave something behind.
      inputs: NBBO close quotes, quote update count
      lookback: 60
      lag: 0
      frame: per symbol-session, trailing
      representation: level and cross-sectional z-score
      failure_mode: >-
        A carried-forward quote reports a stale spread as a live one, so the state
        looks calm exactly where the book has stopped updating.
    - name: microprice
      pattern: microprice_dev*
      role: signal
      hypothesis: >-
        The depth-weighted price leans toward the thin side of the book, so its distance
        from the midpoint is where the next trade is more likely to print (Stoikov, 2018).
      inputs: NBBO close prices and sizes
      lookback: 15
      lag: 0
      frame: per symbol-session, trailing
      representation: level and cross-sectional z-score
      failure_mode: >-
        The deviation is a few tenths of a cent on a wide book and a rounding artifact
        on a one-tick one, so it is not comparable across names until it is ranked.
    - name: volatility
      pattern: r1m*|rv_*|quote_range*
      role: state
      hypothesis: >-
        How much the quoted price is moving is what decides whether an imbalance is
        information or noise, and it varies by name and by hour.
      inputs: NBBO quote midpoint and quote extremes
      lookback: 60
      lag: 0
      frame: per symbol-session, trailing
      representation: level and cross-sectional z-score
      failure_mode: >-
        Every estimator here is a variance over a short window, so a single bad quote
        dominates it and reports turbulence where there was a data error.
    - name: order_flow
      pattern: signed_vol_share*|tick_imb_share*|trade_to_mid*|trades_per_1k_shares*|cross_locked_share*
      role: signal
      hypothesis: >-
        Aggressive volume arrives in runs, so the side that has been paying the spread
        over the last few minutes keeps paying it over the next few.
      inputs: AlgoSeek trade-location and tick buckets, total traded volume
      lookback: 60
      lag: 1
      frame: per symbol-session, trailing
      representation: level and cross-sectional z-score
      failure_mode: >-
        Trade location is assigned against the prevailing quote, so a bar whose quote
        was stale attributes its volume to the wrong side rather than to neither.
    - name: price_impact
      pattern: dollar_vol*|illiq*|kyle_lambda*|trade_range*
      role: state
      hypothesis: >-
        How far a given quantity of order flow moves the price is what separates a
        crowded name from a thin one, and it is the cost side of every signal here.
      inputs: consolidated trade prices, traded dollars across both venues
      lookback: 60
      lag: 1
      frame: per symbol-session, trailing
      representation: level and cross-sectional z-score
      failure_mode: >-
        Amihud is undefined on a bar that printed nothing, and its rolling average
        carries that gap across the whole window rather than skipping the bar.
    - name: hidden_liquidity
      pattern: finra_share_60m*
      role: state
      hypothesis: >-
        The share of volume printing away from the exchanges tracks how much of the
        session's liquidity is not visible in the book, and it moves over hours.
      inputs: FINRA/TRF reported volume, total traded volume
      lookback: 60
      lag: 1
      frame: per symbol-session, trailing
      representation: level and cross-sectional z-score
      failure_mode: >-
        TRF prints are reported with a delay, so the share attributed to a bar is not
        fully public at that bar's close unless it is lagged.
    - name: session_clock
      pattern: time_since_open|time_to_close|is_first_30m|is_last_30m
      role: state
      hypothesis: >-
        Intraday liquidity is U-shaped, so where in the session a bar sits conditions
        every other family without predicting anything on its own.
      inputs: exchange calendar
      lookback: 0
      lag: 0
      frame: per session, from the scheduled open and close
      representation: fraction of the session and block flags
      failure_mode: >-
        Measured against the session's realized bar count rather than its scheduled
        close, it is a quantity no one had at the time.

# 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. It is counted in ROWS OF THE PRICE
# FRAME, and that frame is one-minute: ``config/backtest/base.yaml`` declares
# ``calendar.data_frequency: 1m``, ``load_backtest_prices_for`` returns bars
# measured at 60-second spacing, and ``_compute_rolling_vol`` applies
# ``rolling_std(vol_window)`` to those rows with no resampling. So 520 is 520
# MINUTES - about 1.3 regular sessions - and not the 20 trading days this comment
# claimed until 2026-09-06. The value is unchanged; what was wrong was the
# sentence describing it, which read the decision cadence as the bar size.
# ``periods_per_year`` annualizes Sharpe at daily-equivalent grain (252),
# independently of this window.
#
# Expensive allocators (RP/HRP/MVO_LW) are dropped via expensive_allocators_skip
# below — covariance estimation at 15-min cadence is prohibitive and
# empirically degenerate. MVO_LW carries no per-allocator lookback override
# here for that reason; revisit if expensive allocators are reintroduced
# (a 6-month equivalent at 15-min would need ~3,300 bars).
#
# ``initial_cash`` restored to 1_000_000 (2026-05-16) — the 2026-05-15 SSOT
# migration drop to 100k caused catastrophic degenerate output here
# (avg Sharpe -11 across 339 signal runs; integer-share rounding × the
# nasdaq-100 high-priced tail (BKNG≈$5k, MSTR/NFLX/AVGO $700-1.5k)). 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: 520        # rows of the 1-minute price frame: ~1.3 sessions

mapping:
  class: intraday_rank_and_trade
  position_state_space: long_short
  entry_logic: rank_or_threshold
  sizing: dollar_neutral_or_beta_neutral

costs:
  class: dominant
  # Per-share commission plus measured per-asset half-spread slippage from
  # AlgoSeek NBBO close quotes over the training window. Integer share
  # sizing. The asset_spreads_source parquet is loaded by the backtest
  # cost-preset builder and joined per symbol; default_half_spread_usd is
  # the universe p75 fallback for symbols not in the profile.
  model: per_share_plus_spread
  per_share: 0.0035   # IBKR Pro Tiered top tier (<=300K shares/mo)
  minimum: 0.0
  asset_spreads_source: liquidity_profile.parquet
  asset_spreads_column: median_half_spread_usd
  default_half_spread_usd: 0.025
  spread_convention: half_spread
  source: AlgoSeek minute-bar NBBO close quotes, training window 2020-01-02 to 2021-07-01.
  friction_floor_bps: 5

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:
    # Canonical strategy universe. The 15-min round-trip cost on the
    # cost-expensive tail of the 115-name panel consumes the intraday edge:
    # the featured config collapses out of sample on the full universe
    # (holdout -0.42) but is positive on the cost-feasible subset (the
    # cheapest-to-trade names, frozen per split — see universe.cost_feasible).
    # The full-universe variant is NOT a canonical rank-1 / cohort / DSR
    # candidate; it lives only in the 17_costs.py full-vs-screened comparison,
    # mirroring sp500_options' liquid screen. Read by
    # case_studies.utils.sweep_config.get_universe_filters_for.
    universe_filter: cost_feasible
    # 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
      # top-1 of every pre-cost stage per label. `17_costs` pools {signal, allocation,
      # risk_overlay} - `STAGE_SEQUENCE` minus the stage it writes - because an overlay is a
      # candidate to carry the case study and pricing only its parent would put a cost curve
      # in the chapter for a strategy this case study does not select.
      cost_sensitivity: 1
      risk_overlay: 1           # top-1 of {signal+allocation} per label
    # No expensive allocators on nasdaq. At 15-min frequency the 1.3M-bar
    # panel makes covariance estimation on the rebalance schedule
    # prohibitive; empirically every expensive allocator (risk_parity,
    # mvo_ledoit_wolf, hrp) produces deeply negative Sharpes (-7 to -10)
    # on every nasdaq label and equal_weight ties at the top. The 15-min
    # cadence is a costs-dominated regime where signal-stage selection
    # carries the entire story.
    expensive_allocators_skip: true
    # Ch16 signal-stage selection. Long-short equal-weight top-k on the
    # 15-min schedule across four labels (5m/15m/60m continuous + 15m
    # direction). The classification label fwd_dir_15m is a first-class
    # signal alongside the continuous labels and carries the same top_k
    # grid.
    top_k_grid:
      fwd_ret_5m:  [5, 10, 20]
      fwd_ret_15m: [5, 10, 20]
      fwd_ret_60m: [5, 10, 20]
      fwd_dir_15m: [5, 10, 20]
    # Ch17 portfolio: reuses top_k_grid above and sweeps over allocators.
    # Nasdaq runs only the three fast allocators — the moment-based ones are
    # too slow at intraday frequency (``expensive_allocators_skip: true``).
    # No max_weight cap on inverse_vol — see
    # memory/feedback_max_weight_caps_intentionally_absent.md (would force
    # equal-weight at top_k=5 and defeat the moment-allocator comparison).
    allocators:
      - {name: equal_weight,    method: equal_weight}
      - {name: score_weighted,  method: score_weighted}
      - {name: inverse_vol,     method: inverse_vol}
    # Ch18 cost sensitivity (bps regime).
    cost_grid_bps: [0, 1, 2, 3, 5, 7, 10, 15, 20, 30, 50]
    # Companion per-share cost regime: dollars per share half-spread. Swept
    # alongside cost_grid_bps for CSes whose declared cost model is
    # per_share_plus_spread. 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]
    # Alternative rebalance cadences explored by Ch18's cadence × cost
    # heatmap (17_costs.py Section 4). The first entry must be the CS's
    # default cadence (``decision.bar_frequency``). Each cadence reuses the
    # ``cost_grid_half_spread_usd`` grid above. Tokens are the same cadence
    # vocabulary recognized by the engine; the notebook pulls the list from
    # here, never hardcoded.
    cadence_sweep: [15_minute, 30_minute, 1_hour, 4_hour]
    # nasdaq100 v4 slot-mechanism sweep (Ch16 signal stage replacement).
    #
    # Adds two selection methods to the canonical Ch16 expansion:
    #   - slot_persistent_signal_exit: per-symbol rolling-percentile entry,
    #     fixed weight-per-slot allocation (slots ARE the allocation —
    #     Ch17 allocator stage is skipped for these rows), max-hold +
    #     optional signal-exit (rolling lower quantile crossover) for
    #     position lifecycle. See case_studies/utils/slot_strategy.py for
    #     the mechanism. Sub-params live under selection_method_config in
    #     the flattened scheme dict.
    #   - eq_w_topk: the canonical equal-weight top-K baseline, kept as
    #     a within-sweep comparison. Direction grid covers long_only,
    #     long_short.
    #
    # slot × long_short is intentionally dropped — slot books are
    # single-direction by construction (one symbol cannot occupy a long
    # AND short slot simultaneously). The cross-asset long-short story
    # belongs to eq_w_topk / quintile_long_short.
    #
    # Per-prediction cardinality:
    #   slot:      long_q(3) × max_slots(3) × hold_bars(2) × exit_signal_q(2)
    #              × bars_per_day_grid(1) = 36 (long_only only)
    #   eq_w_topk: top_k(3) × direction(2) = 6
    # 42 per prediction, and `signal_passes` below says how many predictions
    # each of them runs on.
    #
    # `hold_bars` and `exit_signal_q` were cut here, from 3 and 4 values to 2
    # each, on 2026-09-11. Both keep their endpoints: hold_bars keeps 2h and 8h
    # and drops the 4h midpoint, exit_signal_q keeps "off" and the median
    # crossover and drops 0.30 and 0.70. A midpoint is what a grid can lose
    # without losing the comparison it draws - the question the sweep asks is
    # whether a longer hold and a signal exit move the result, and two values
    # answer it. `long_q` and `max_slots` are untouched: `14_backtest.py`'s
    # Act 2 selects the carrier by `slots == 10` and `entry_q == 0.9`
    # literally, so removing either value empties the table that section
    # prints.
    signal_nasdaq100:
      selection_method: [slot_persistent_signal_exit, eq_w_topk]
      long_q: [0.90, 0.95, 0.99]
      direction: [long_only, long_short]
      max_slots: [5, 10, 20]
      # hold_bars and bars_per_day count rows of the grid the slot simulator walks, which is
      # the PRICE grid, now one minute. Restated from the fifteen-minute counts they carried
      # (8/16/32 and 14) at the same wall-clock durations, so the sweep asks the same question
      # of the same intervals rather than silently asking about 8, 16 and 32 minutes.
      hold_bars: [120, 480]               # 2h, 8h, in minute bars
      exit_signal_q: [null, 0.50]         # null disables signal-exit
      pred_freshness_max_min: 14          # tolerated staleness of a prediction, in minutes
      bars_per_day_grid: [210]            # was 14 bars of 15 minutes; same 3.5h window
      top_k_grid: [5, 10, 20]
      lookback_days: 21
    # How the signal stage walks the grid above. Two passes, because the whole
    # grid over every prediction is not affordable here and nothing else in the
    # book has this shape: every other case study declares 2 to 4 entry schemes
    # and one universe, and no completed case study holds more than about 3,000
    # signal backtests.
    #
    # Counted against this registry and measured on 2026-09-11. A backtest of
    # this panel costs 31.8 s (four arms timed on the live 10.1M-row price
    # frame, both universes, 29.2 s to 36.3 s), and the registry offers 741
    # admissible prediction sets: 226 for each of the three continuous labels
    # and 63 for `fwd_dir_15m`.
    #
    #   pass 1  baseline    741 × 3 canonical arms  × cost_feasible = 2,223
    #   pass 2  mechanism    32 × 42 arms           × cost_feasible = 1,344
    #   pass 2  reference    32 × 3 canonical arms  × full          =    96
    #                                                       total   = 3,663  ≈ 32 h
    #
    # Crossing all 117 arms with both universes over all 741 predictions, which
    # is what this block replaces, is 173,394 backtests and about 1,500 hours.
    # The prediction count is not what was wrong with it: 741 sets is what three
    # DL families at 20 epoch checkpoints, 15 gbm configs at 10 tree-count
    # checkpoints and 16 linear configs come to, and every one of them is a
    # configuration this case study fitted on purpose. The arm count was.
    signal_passes:
      # Pass 1 ranks every admissible prediction on three equal-weight
      # concentrations, on the universe the carrier actually trades. Ranking on
      # the full universe instead would order the predictions by a Sharpe no
      # selected strategy ever earns, because the cost-expensive tail of the
      # panel is exactly what `universe.cost_feasible` removes.
      baseline_schemes: [ew_top5, ew_top10, ew_top20]
      baseline_universe: cost_feasible
      # Pass 2 runs the mechanism grid on the strongest predictions of pass 1,
      # per label, ranked by pass-1 validation Sharpe.
      #
      # By Sharpe and not by IC. `top_n_predictions.signal` is the other lever
      # that would cut this grid and it is the wrong one: `load_prediction_index`
      # ends `ORDER BY m.ic_mean DESC NULLS LAST`, so any positive value there
      # screens the signal stage on rank correlation and decides which models are
      # backtested at all before a backtest has run. All nine case studies
      # declare `signal: 0` for that reason, and it stays 0 here.
      mechanism_top_n: 8
      # The full universe is kept as a reference arm rather than a second copy of
      # the sweep: the canonical concentrations only, on the pass-2 predictions.
      # That is what `17_costs` reads for its full-vs-screened comparison - it
      # prices the top-1 of each pre-cost stage, which is always inside pass 2 -
      # and what Act 1 of `14_backtest` reads to show the unscreened baseline.
      reference_schemes: [ew_top5, ew_top10, ew_top20]
      reference_universe: full
    # 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}

# The mean-forecast ensemble this case study features, and the only thing in the repository
# that produces `family: ensemble` rows. Declared rather than assembled in the notebook
# because the member set is what the ensemble *is*: change the rule and it is a different
# object under the same name.
#
# The lesson this case study ends on is that choosing among a family's configurations on a
# validation window does not carry to a later window, while the family itself does. An object
# that makes no choice is what shows it, and that object has to be registered like a model so
# the backtest stage, the cohort comparison and the holdout all read it the way they read one.
ensemble:
  # The family averaged, and the rule that decides which of its configurations are in.
  #
  # 31 leaves is the boundary between the regularized presets and the one that is not:
  # `leaves_7`, `leaves_15` and `leaves_31` carry `lambda_l1`, `lambda_l2`, `bagging_fraction`
  # and `feature_fraction`; `default` carries none of them and trains at LightGBM's own
  # default of 31 leaves; `leaves_63` is the only configuration above the line. The rule keeps
  # twelve of the fifteen for each continuous label - three losses across four leaf counts.
  #
  # A count is not declared here on purpose. The rule is the declaration and the members are
  # resolved from what is registered, so a configuration added to `config/training/*.yaml`
  # joins the ensemble and a number written down here could only disagree with it.
  member_family: gbm
  max_num_leaves: 31
  method: mean_forecast
  config_name: gbm_mean_leaves31
  # Each member enters at its last checkpoint. Taking the checkpoint with the best validation
  # metric would select on the window the ensemble is then measured on, which is the thing
  # the ensemble exists not to do.
  checkpoint: last
  # The entry scheme the ensemble is backtested on, by the name `get_entry_schemes_for`
  # gives it: 10 slots, an 8-hour maximum hold, a 0.90 entry quantile, no signal exit, on the
  # 210-bar execution window. `14_backtest.py` Act 2 reads the carrier by `slots == 10` and
  # `entry_q == 0.9` literally, so this arm and that filter have to agree.
  #
  # One arm and not the grid: the ensemble is not a sweep participant. It is the object the
  # chapter carries, backtested on the design the sweep already chose, and putting it through
  # the sweep would make it one more configuration to select among.
  featured_scheme: slot_l_lq90_s10_h480_noexit_b210
  universe_filter: cost_feasible

# The model specification 04_model_based_features estimates under. Declared here rather than
# typed into the notebook so that a window is one number in one place, readable without opening
# the code that consumes it.
#
# This case study fits no GARCH, no regime chain and no ARIMA: its three procedures are a
# heterogeneous autoregression of realized variance, a rolling Fourier transform and a rolling
# depth-2 path signature. Only the first estimates anything, which is why only `har` carries a
# refit schedule.
model_based:
  har:
    # The three lengths of history the regression averages squared one-minute returns over,
    # in bars. Corsi's day/week/month on a minute grid: minutes, a quarter hour, an hour.
    components: [5, 15, 60]
    # Bars of history each refit reads. Two hours - long enough for four coefficients to be
    # identified, short enough that the fit follows the day rather than the quarter. It is
    # also the burn-in: the first `fit_window + 1` bars of a symbol carry no HAR value,
    # because they are what pays for the first estimate.
    fit_window: 120
    # Bars between refits. One, which is the finest cadence the panel has and the reason this
    # artifact carries no fold column: the coefficients describing bar t come from a
    # regression whose last observation is bar t-1, so no fold boundary enters the fit and a
    # bar's value is the same value whichever fold selects it.
    refit_every: 1
    # Usable observations a refit window needs before it is fitted at all. Below this the four
    # coefficients are estimated off too few points to identify them, and the bar keeps a null
    # rather than a fit on whatever survived the warm-up.
    min_train_obs: 20
  spectrum:
    # Bars each trailing window spans. Sixty resolves repetition periods up to an hour, which
    # is the longest intraday rhythm this panel is asked about.
    window: 60
    # The period, in bars, above which power counts as low frequency. Twenty separates a slow
    # drift in activity from minute-to-minute churn, and it is what `*_low_freq_ratio` is the
    # share of.
    low_frequency_period: 20
  signature:
    # Bars each signature path spans. Thirty is the horizon over which order flow and price are
    # expected to lead one another, which is the quantity the cross terms measure.
    window: 30

evaluation:
  n_splits: 2
  train_size: 6M
  val_size: 6M
  holdout_start: '2021-07-01'
  holdout_end: '2021-12-31'
  calendar: NYSE
  periods_per_year: 252  # NYSE 5d/wk (daily MTM despite intraday signal cadence)

labels:
  primary: fwd_ret_15m
  # The outcome horizon each label's name states. Declared rather than left to the default,
  # because `resolve_label_horizon` falls back to the BUFFER when no horizon is given - so
  # widening a buffer below would silently lengthen the return the label measures, and
  # `fwd_ret_15m` would be a sixteen-minute return that every consumer still called fifteen.
  horizons:
    fwd_ret_15m: 15min
    fwd_ret_5m: 5min
    fwd_ret_60m: 60min
    fwd_dir_15m: 15min
  # The purge, which is the horizon PLUS ONE BAR - wider than the horizon, and deliberately.
  # The entry leg is the VWAP of the bar after the decision and the exit leg is H past the
  # entry, so a label at t consumes a quote at t+H+1: one bar past the horizon its name
  # states. A buffer of H would leave the last bar of every training window inside the first
  # validation label's window. The extra minute is not conservatism; it is the width of the
  # window that the horizon alone does not describe.
  buffer: 16min
  variants:
    - fwd_ret_5m
    - fwd_ret_60m
    - fwd_dir_15m
  variant_buffers:
    fwd_ret_5m: 6min
    fwd_ret_60m: 61min
    fwd_dir_15m: 16min
  # How many slots of the DECISION schedule to advance per trade, so a position is held for
  # its label's horizon and no two holdings overlap. A slot is one interval of the declared
  # cadence, so these are `ceil(horizon / cadence)` and nothing else.
  #
  # That reading is only true because `resolve_rebalance_timestamps` now builds the schedule
  # from the declared cadence. It used to return the panel's own timestamps unchanged for
  # every intraday cadence, and this panel carries every minute, so these values thinned
  # minutes rather than fifteen-minute slots and three of the four labels traded every
  # minute (ml4t/agent-workspace#187). The values did not change; what they count did.
  # Every label now trades on a cadence equal to its own horizon (decision.cadence_by_label
  # above), so one slot is one holding period and no label needs a multi-slot step. The
  # `fwd_ret_60m: 4` this replaces was `ceil(60 / 15)` against a single fifteen-minute
  # default; the cadence carries that now, and carries it for the Ch18 sweep arms too,
  # which differ from the default in exactly this token.
  rebalance_step:
    fwd_ret_15m: 1
    fwd_ret_5m: 1
    fwd_ret_60m: 1
    fwd_dir_15m: 1
  # Continuous return that each classification label is derived from.
  classification_eval_label:
    fwd_dir_15m: fwd_ret_15m

modeling:
  dl:
    # The backend the sequence families are trained on in production. Declared rather
    # than defaulted: the four deep-learning notebooks refuse to train on a device other
    # than the one asked for, because the device is part of what the run is registered
    # under, and a notebook that invents `gpu` from a missing section makes a
    # CUDA-free environment fail on a requirement nobody wrote down. A run on other
    # hardware sets the notebook's DEVICE parameter, which is then what gets registered.
    device: gpu
    # How far apart the training windows of one sequence configuration sit, counted in
    # label horizons. This is the only case study that needs the key: every row of a panel
    # starts a window, and this panel is minute bars, so an unspaced fold builds about 4.0
    # million near-identical overlapping sequences where the daily case studies build tens
    # of thousands. Measured on the canonical plan, 2026-09-02: fold 0 has 4,066,041
    # training rows and fold 1 has 3,997,571, across 115 symbols.
    #
    # One window per label horizon. Consecutive windows of a symbol then carry labels that
    # do not overlap: the window ending at t is scored on the return from t to t+H, and the
    # next one starts at t+H. Every window is still an example the model has to fit; what
    # changes is that no two of them are the same example seen twice.
    #
    # Declared in horizons rather than in windows because the horizon differs by label, and
    # one line then covers all four: `fwd_ret_15m` and `fwd_dir_15m` stride 15 observations,
    # `fwd_ret_5m` strides 5 and `fwd_ret_60m` strides 60, each spacing its own labels
    # exactly. A single count could only be right for one of them. On fold 0 the primary
    # label draws about 271,000 windows, and the count is derived at run time and printed,
    # not written down here - it differs between folds because the folds differ in length.
    #
    # This replaces `max_train_sequences: 750000`, which had been carried unchanged since
    # the original release and through the stage-06 rebuild that moved the folds underneath
    # it, so nothing derived it. At 750,000 a window was drawn every 5.4 minutes, finer than
    # the 15-minute horizon, so neighbouring training windows carried overlapping labels and
    # the extra 2.8x of them were re-presentations of examples already in the batch. That is
    # not wrong - correlated examples are not invalid ones - but it was a number with no
    # argument behind it, and the number is part of the training identity
    # (ml4t/agent-workspace#1015).
    #
    # `max_train_sequences` is still available and is what a case study declares when it
    # wants the overlap: it fixes the count and lets the spacing follow. The two are
    # mutually exclusive.
    train_sequence_stride_horizons: 1
  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

causal:
  treatment: signed_vol_share
  # Bars the treatment's own construction window spans, which is what the placebo block has
  # to cover: permuting signed_vol_share 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:
  #
  # `signed / TRADED_VOLUME` in 03_financial_features, both from the same bar. The smoothed
  # versions this feeds are separate columns; the treatment itself spans one bar.
  treatment_window: 1
  confounders: [rel_spread_close, rv_5m, r1m]
  method: walk_forward_dml

```