Project 4 Reader Guide: From Fields to Signals and Execution

The question to answer first

Value, momentum, and quality fields differ in units and coverage. First ask whether each input was known at signal time. Then preprocess and combine under same-date policies, and finally define the support on which positions may exist. Signal targets, execution-filtered targets, and realized holdings are three objects. Neutral targets do not guarantee neutrality after fills.

Read the project objective, signal method, event timeline, seven-field contract, dictionary, fixed rules, real-input gaps, and source feasibility in order. Chapters 24–26 define the economics; Chapter 28 derives residuals, scaling, combination, and constraints. Chapter 27 carry/BAB methods are not in the current seven-field contract. After the signal layer, continue with Chapter 29, the ledger-interface guide, and the cost lab; consult its data dictionary, research report, and complete six-question solutions as needed. The cost lab’s Exercise 1–6 and answers remain separate; this BIL15B candidate adds new constructed P4-Txx integrative exercises and complete solutions.

Default runs and user-supplied inputs

The accepted I 3 shared project reader instructions govern the reader kernel, clean-kernel execution, package-root discovery, and new RUN_DIR. This guide inherits that full offline/user-supplied-input boundary. Public availability does not establish permission to download, use, or redistribute. A real-data branch requires separate rights and execution authorization.

The signal Notebook's Chinese Markdown projection and English Markdown projection preserve I 3's 19 cells, 9 code cells, and code/output/metadata identity. Their present logical locations support language review. A final standalone reader package must place the selected language at the default project path and rebind package-root discovery and relative links. Current teaching review does not establish a standalone-package cold start.

The default uses seed 20261210, 240 companies including 40 banks, and 96 months, with no real company data. Its first cell creates runs/p04_signal_layer_<uuid>/; a new image goes to figures/a5/p04_signal_layer.png within that directory. The runner separately saves an execution copy/status. The default signal path uses no French data and does not run historical CLI/pytest entry points. Markdown descriptions of historical results do not mean that these tables are regenerated.

Check the current signal method in order

1. Form support from listing status known at signal time, with at least 60 listed names. Require both filing availability and ingestion before the cutoff, and select an entire filing version together.
2. Preserve economic direction in the seven fields. Nonpositive denominators and inapplicable bank measures become missing, not zero.
3. Same-date 1%/99% winsorization needs 30 values. Industry/size OLS uses a common complete-pair set and at least 2 residual degrees of freedom. Sample z-scores need 10 values.
4. Each style needs at least 1 available nonzero-weight field; the composite needs at least 2 styles. Renormalize over available terms and preserve missingness when a gate fails.
5. Alpha validity determines only the initial quintile ranking. Optional industry projection uses all listed support coordinates, allowing listed names without alpha to hedge. Off-support weights are exactly zero.
6. For a nonzero projection satisfying homogeneous industry/net-zero constraints, use one positive scalar to restore gross 2. No long–short direction or a degenerate projection means cash. A nonzero constraint target cannot use this scaling argument.
7. Report industry/net/gross, support residuals, coverage, and missingness separately. Post-execution information belongs to the ledger and diagnostics and cannot rewrite signals.

Read three types of results separately

Historical A5/A6 constructed results retain the old full-column projection and negative findings. C13 corrects the support/gross/60-name protocol; see the current note. The default Notebook uses the current constructed method and cannot turn old tables into a new run. None of these establishes real-market alpha, universal absence of leakage, or tradable capacity.

Cost materials separately identify historical R2R6 and the 2026-10-14 current-signal summary. The English cost Notebook and its Chinese projection start from frozen I3: fixed-parent-order and partial/zero-fill examples are computed afresh; the other walk-forward, sensitivity, ablation, and ledger-audit displays replay the accepted summary and are not reruns. The base run_backtest raises on a nonzero target that is already delisted when it reaches execution. The A6 wrapper zeros such targets in an execution copy, without rewriting the signal or rescaling other names. The cost walk-forward consumer instead checks listing status on each execution date and processes existing delisted holdings first. These are separate execution paths. Complete bilingual cost/execution/backtest teaching continues through Chapters 29–31, with consumer review in Chapters 35–36; full P4 teaching still depends on BIL15B validation acceptance, and real-market research remains incomplete.

Real PIT statements, total returns, historical membership, capitalization, industry versions, trading calendars, financing/borrowing, and capacity gaps remain. This BIL15B candidate now includes new P4 integrative exercises and complete solutions, with independent P4-Txx IDs. The reading route is signal layer → Chapter 29 ledger → cost lab → Chapter 30 bias → Chapter 31 validation → these constructed cases. Cost Exercise 1–6 remain separate. The candidate counts toward full P4 teaching only after the root independently accepts BIL15B, alongside the accepted BIL14 and BIL15A dependencies. BIL17 consumer review and BIL23 English Notebook, portable assembly, cold start, and global-link checks remain. Real-market research remains planned pending PIT data and rights, borrowing, a trading calendar, and capacity evidence.