mirrors / stefan-jansen--machine-learning-for-trading

项目文件夹

stefan-jansen--machine-learning-for-trading
文件
main
T
wehub-resource-sync 0249a267e1 chore: import upstream snapshot with attribution
2026-07-13 13:26:28 +08:00

469 行
14 KiB
YAML
原始文件 永久链接 Blame 文件历史

此文件含有模棱两可的 Unicode 字符
此文件含有可能会与其他字符混淆的 Unicode 字符。 如果您是想特意这样的，可以安全地忽略该警告。 使用 Escape 按钮显示他们。
strategy_id: nasdaq100_microstructure
setup_version: v1
universe:
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
- TEAM
- TMUS
- TSLA
- TTWO
- TXN
- UAL
- ULTA
- VRSK
- VRSN
- VRTX
- WBA
- WDAY
- WDC
- WLTW
- XEL
- XLNX
- ZM
n_assets: 114
eligibility_rule: nasdaq100_membership
# 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 114 universe collapses out of sample (featured
# config holdout -0.42) while the cost-feasible universe is positive; the
# screen is load-bearing. See 16_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:
bar_frequency: 15_minute
decision_snapshot: bar_close
execution_delay: 1_bar
execution_price_assumption: next_bar_open_or_vwap
# 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. 15-min bars (26 bars/day);
# 520 bars ≈ 20 trading days ≈ 1 month of intraday coverage. ``periods_per_year``
# annualizes Sharpe at daily-equivalent grain (252); allocator windows are
# measured in raw 15-min bars regardless.
#
# 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 # ~20 trading days of 15-min bars
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 114-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 16_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_cheap: 0 # score_weighted, inverse_vol — all signal preds
allocation_expensive: 0 # nasdaq has no expensive allocators (see expensive_allocators below)
cost_sensitivity: 1 # top-1 of {signal+allocation} per label
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
expensive_allocators: []
# 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.
# Cheap tier only on nasdaq (see ``expensive_allocators`` above).
# 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 (16_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(3) × exit_signal_q(4)
# × bars_per_day_grid(1) = 108 (long_only only)
# eq_w_topk: top_k(3) × direction(2) = 6
# ~114 per prediction × ~12 representative preds ≈ ~1,400 backtests
# for the signal stage on val (less than the original 2,800 estimate
# because slot × long_short is dropped).
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: [8, 16, 32] # 2h, 4h, 8h at 15-min grid
exit_signal_q: [null, 0.30, 0.50, 0.70] # null disables signal-exit
pred_freshness_max_min: 14 # variant-A: freshest 1m pred per 15m bar
bars_per_day_grid: [14] # 15m only; 1m cadence failed comprehensively
top_k_grid: [5, 10, 20]
lookback_days: 21
# 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: 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
buffer: 15min
variants:
- fwd_ret_5m
- fwd_ret_60m
- fwd_dir_15m
variant_buffers:
fwd_ret_5m: 5min
fwd_ret_60m: 60min
fwd_dir_15m: 15min
# Vectorized-backtest thinning step per label: number of schedule slots
# to advance per trade so holding periods don't overlap.
rebalance_step:
fwd_ret_15m: 1
fwd_ret_5m: 1
fwd_ret_60m: 4 # 15min schedule, 60min horizon -> ceil(60/15) = 4
fwd_dir_15m: 1
# Continuous return that each classification label is derived from.
classification_eval_label:
fwd_dir_15m: fwd_ret_15m
modeling:
gbm:
libraries: [lightgbm]
preset: default
device: cpu
causal:
treatment: signed_vol_share
confounders: [rel_spread_close, rv_5m, r1m]
method: walk_forward_dml
在新工单中引用 查看 Git Blame 复制永久链接