I built an end-to-end U.S. large-cap absolute-return research system combining cross-sectional modelling, purged walk-forward evaluation, constrained portfolio construction, regime-aware risk control, next-open execution, and sleeve-level P&L attribution: a market-neutral Alpha Core plus a separately governed Directional Risk Overlay.
The complete strategy performed well during development and still earned a positive return in the locked holdout, now observed. I did not stop at headline performance: I traced the holdout P&L back to the two sleeves and costs. That attribution changed the conclusion: the combined portfolio performed well, but the gain was overlay-led; independent stock-selection alpha remains unvalidated.
All results are simulated. The study uses a fixed, retrospectively curated universe of 174 U.S. large-cap equities rather than historical point-in-time index membership. The holdout has only 76 observations; annualized metrics are unstable over this short window and are not long-run estimates.
- Build the complete strategy. Connect cross-sectional signals to a dollar-, sector-, and estimated-beta-controlled Alpha Core, then govern directional exposure with a separate overlay that leaves stock rankings unchanged.
- Establish development performance. Over 2020-09-01 to 2026-03-26, the combined portfolio delivered 23.49% CAGR, 1.24 Sharpe, and -16.10% MDD after 5 bps one-way costs. The Alpha Core alone had 0.94 Sharpe and 0.0059 realized beta to SPY.
- Observe the locked holdout. With the design frozen, 2026-04-01 to 2026-07-21 produced +7.71% total return, 1.39 annualized Sharpe, and -8.23% MDD at 5 bps across 76 observations. Both 23.49% CAGR and +7.71% total return describe the complete portfolio, not pure stock selection.
- Look beyond headline performance. Holdout attribution showed Alpha Core -1.58% gross, Overlay +9.68% gross, and transaction costs about -0.14%. These arithmetic contributions sum to +7.96%; compounding daily net returns gives the reported +7.71%.
- Revise the claim. Positive combined return, but overlay-led. The project's contribution is the deliberate investigation of an apparently successful result, identification of its return source, and revision of the conclusion. A positive holdout does not establish stable, independent stock-selection alpha.
174 equities · 19 model inputs · 1,398 development sessions ·
4 transaction-cost assumptions · 76-session holdout
| Evidence | Reviewed result |
|---|---|
| Development, 5 bps one-way cost | 23.49% CAGR · 1.24 Sharpe · -16.10% maximum drawdown |
| Development cost stress, 20 bps | 21.57% CAGR · 1.16 Sharpe |
| Market-neutral Alpha Core, development | 0.94 Sharpe · 0.0059 realized beta to SPY |
| One-time holdout, 76 sessions | +7.71% · 1.39 annualized Sharpe · -8.23% maximum drawdown |
| Holdout sleeve attribution | Alpha Core -1.58% gross · Directional Risk Overlay +9.68% gross · costs about -0.14% |
The combined portfolio result and the stock-selection result are not the same claim. The development period supports the fixed system specification; the short holdout supports the combined portfolio only over that later window; the holdout attribution leaves independent stock-selection alpha unvalidated.
| Evidence layer | Question it answers | Role in the conclusion |
|---|---|---|
| Development | What did the fixed research design produce over the main sample? | Establishes the primary performance and risk profile |
| Cost sensitivity and robustness | Does the result disappear under nearby assumptions or samples? | Tests dependence on one convenient specification |
| One-time holdout | What happened after the design was fixed? | Adds later, non-development observations |
| Sleeve attribution | Which portfolio decision generated the result? | Separates combined performance from component stability |
flowchart LR
Q["Research question"]
T["Tradable definition<br/>information · label · execution"]
M["Walk-forward model<br/>two cross-sectional tails"]
P["Portfolio system<br/>Alpha Core + directional overlay"]
V["Evidence protocol<br/>costs · robustness · holdout"]
A["Attribution<br/>what drove the result?"]
N["Next research question<br/>Alpha Core stability"]
Q --> T --> M --> P --> V --> A --> N
The sections below follow this chain from problem definition to evidence and interpretation.
A cross-sectional model can rank securities correctly while the resulting portfolio still earns most of its return from a broad market exposure. For an absolute-return objective, model accuracy, portfolio neutrality, directional risk, and implementation cost must therefore be treated as connected but distinct research problems.
The project asks three linked questions:
- Can upper- and lower-tail stock outcomes be estimated using only information available at each decision time?
- Can those estimates be converted into a market-neutral stock-selection sleeve after sector, beta, concentration, and turnover controls?
- Can directional exposure be governed separately so that total portfolio P&L remains attributable to the decision that created it?
The portfolio is represented as two separately governed sleeves:
combined portfolio weights
= market-neutral Alpha Core weights
+ Directional Risk Overlay weights
The Alpha Core expresses relative stock selection. The Directional Risk Overlay changes aggregate exposure in response to market state but leaves the cross-sectional ranking unchanged.
The same separation is maintained in the return ledger:
combined net P&L
= Alpha Core gross P&L
+ Directional Risk Overlay gross P&L
- transaction costs
This decomposition shapes the data timing, model evaluation, portfolio checks, and final interpretation. A positive combined return cannot automatically be described as standalone stock-selection alpha.
flowchart TB
subgraph I["Information and data"]
direction LR
I1["174-stock OHLCV panel"]
I2["SPY · VIX · macro proxies · headline activity"]
I3["schema · coverage · price · tradeability checks"]
I1 --> I3
I2 --> I3
end
subgraph R["Cross-sectional research"]
direction LR
R1["19 model inputs"]
R2["open t+1 to open t+11 target"]
R3["upper-tail and lower-tail classifiers"]
R4["rolling training · scheduled refit · purge"]
R1 --> R2 --> R3 --> R4
end
subgraph P["Portfolio construction"]
direction LR
P1["cross-sectional scores"]
P2["inverse-volatility initialization"]
P3["dollar · sector · beta controls"]
P4["Directional Risk Overlay"]
P5["gross · net · name limits"]
P1 --> P2 --> P3 --> P4 --> P5
end
subgraph E["Execution and evidence"]
direction LR
E1["next-open weights"]
E2["open-to-open P&L"]
E3["turnover · costs · attribution"]
E4["robustness · holdout · release"]
E1 --> E2 --> E3 --> E4
end
I3 --> R1
R4 --> P1
P5 --> E1
E4 -. "component diagnosis" .-> R1
The result is the output of connected research contracts rather than a single model score. A valid run requires agreement across the information, modelling, portfolio, execution, accounting, and evidence layers.
The research panel aligns constituent OHLCV with market and contextual inputs.
| Input group | Role in the system |
|---|---|
| Constituent OHLCV | Stock-level features, volatility, liquidity, labels, and tradable returns |
| SPY | Market return, trend, volatility, and beta reference |
| VIX | Market stress and overlay risk scaling |
| Rates and credit proxies | Macro and risk-condition context |
| Sector proxy | Cross-sectional context and exposure diagnostics |
| Headline activity | Event-intensity context without publishing licensed text |
Before model evaluation, the full implementation checks schema, duplicate keys, price bounds, coverage, tradeability, date alignment, and file identity. The public reference implementation exposes the cross-sectional schema, uniqueness, finite-value, and information-time contracts using synthetic data.
The fixed universe is an explicit limitation. It makes the experiment reproducible but introduces survivorship and membership-selection bias; no claim is made that the universe reproduces historical index membership.
The model matrix contains 19 inputs: 11 stock-factor families plus market, rates, credit, sector-proxy, and headline-activity variables. Exact feature definitions and transformations are not published.
Two horizons must remain distinct:
decision and execution
features observed after close t
-> portfolio decision formed from the close-t information set
-> execution at open t+1
-> daily P&L from open t+1 to open t+2
model target
decision at t
-> forward return from open t+1 to open t+11
The longer prediction target creates unresolved labels near every evaluation boundary. Training therefore ends early enough for all training labels to resolve before testing begins. The effective separation is 11 sessions, and an overlapping split is treated as invalid rather than merely discouraged.
The modelling stage treats the upper and lower cross-sectional tails as different classification problems. Each side combines LightGBM with an L2-regularized logistic model. This retains nonlinear interactions while providing a structurally different model component.
observable feature panel
-> upper-tail probability
-> lower-tail probability
-> combined relative-opportunity score
-> daily cross-sectional ranking
The dual-tail formulation avoids assuming that the mechanisms associated with potential winners and potential losers must be symmetric.
| Design element | Fixed research choice |
|---|---|
| Prediction problem | Separate classification of upper and lower return tails |
| Executable target | Open t+1 to open t+11 forward return |
| Model combination | Deterministic LightGBM/L2-logistic ensemble for each tail |
| Training history | Rolling 1.5-year window |
| Refit schedule | Every 63 sessions |
| Label separation | Effective 11-session purge |
| Score schedule | Refreshed on every feature date |
| Score stabilization | Exponentially weighted smoothing |
The refit schedule and score schedule are intentionally different. Models are refitted periodically, while daily cross-sectional scores are refreshed using the most recent fitted state and currently available features.
Each evaluation step therefore follows the same order:
resolve eligible training labels
-> enforce purge before the test boundary
-> fit or reuse the scheduled model state
-> score the current cross-section
-> pass scores to portfolio construction
The stock-selection sleeve is built as a sequence of portfolio operations:
upper-tail probability and lower-tail probability
-> combined cross-sectional ranking
-> long and short opportunity sets
-> inverse-volatility starting weights
-> dollar-neutral projection
-> sector-neutral projection
-> estimated-beta-neutral projection
-> gross and single-name controls
-> scheduled rebalance and target smoothing
Neutrality is evaluated before the directional sleeve is applied. This keeps a failure in stock-selection portfolio construction from being hidden by the overlay.
| Portfolio diagnostic | Purpose |
|---|---|
| Net exposure | Detect unintended directional exposure in the Alpha Core |
| Gross exposure | Control total long-plus-short capital usage |
| Sector exposure | Prevent the ranking from becoming a sector allocation |
| Estimated beta exposure | Reduce dependence on broad market direction |
| Single-name exposure | Limit concentration in one security |
| Turnover | Connect target changes to implementation cost |
The Directional Risk Overlay is formed only after the Alpha Core has passed its own exposure checks. SPY trend, realized volatility, and VIX govern aggregate net exposure. The overlay does not rerank stocks and does not directly trade a market ETF or futures contract. It adds exposure across securities already in the Alpha Core long book, subject to final gross, net, and single-name limits.
This structure allows the two decisions to be governed and measured separately, but it is not a pure statistical alpha-versus-beta factor decomposition. The term “sleeve-level P&L attribution” is used because it matches the actual portfolio construction.
The execution convention is designed together with the label rather than added after modelling:
information available after close t
-> target weights for open t+1
-> realized portfolio return from open t+1 to open t+2
One-way turnover is defined as half the cross-sectional absolute change in weights:
turnover_t = 0.5 × Σ |weight_t - weight_(t-1)|
cost_t = turnover_t × one_way_cost_bps / 10,000
Every period must satisfy the accounting identity:
Alpha Core gross P&L
+ Directional Risk Overlay gross P&L
- transaction costs
= combined net P&L
The same realized turnover is evaluated at 0, 5, 10, and 20 bps. This isolates cost sensitivity from changes in model or portfolio decisions.
Development, robustness, and holdout evidence have different roles. The development period is used to establish the specification; robustness checks challenge its dependence on nearby choices; the holdout is evaluated only after the research design is fixed.
| Failure mode | System control | Observable output |
|---|---|---|
| Future information enters a decision | Timestamped information set | Decision-clock validation |
| A forward label overlaps evaluation | Label resolution plus purge | Walk-forward split validation |
| Model output drops part of the universe | Full cross-sectional coverage check | Score-schema failure |
| Alpha Core carries directional exposure | Dollar and estimated-beta checks | Exposure report |
| Ranking becomes sector allocation | Sector neutrality check | Per-sector exposure report |
| Overlay changes security selection | Overlay restricted to existing long names | Separate sleeve weights |
| Trading assumptions and returns use different intervals | Shared next-open convention | Execution audit |
| Reported return cannot be reconstructed | Period-level sleeve and cost ledger | Exact reconciliation gap |
| A result depends on one cost assumption | Fixed turnover under four cost levels | Cost-sensitivity curve |
| Holdout alters the chosen design | Specification fixed before evaluation | Separate holdout evidence |
The primary development window spans 2020-09-01 through 2026-03-26.
| One-way cost | CAGR | Annualized Sharpe | Maximum drawdown | Total return |
|---|---|---|---|---|
| 0 bps | 24.14% | 1.27 | -15.68% | 231.89% |
| 5 bps | 23.49% | 1.24 | -16.10% | 222.39% |
| 10 bps | 22.85% | 1.21 | -16.73% | 213.17% |
| 20 bps | 21.57% | 1.16 | -17.97% | 195.50% |
At the primary 5 bps assumption, 90.06% of available rolling 252-session returns were positive and four of five complete calendar years were profitable. The separately measured Alpha Core produced a 0.94 Sharpe and 0.0059 realized beta to the supplied SPY reference.
Robustness checks ask whether the development result depends on a single convenient specification. They reduce selection risk but do not eliminate sampling uncertainty.
| Dimension varied | Reviewed outcome |
|---|---|
| One-way transaction cost | Sharpe remained 1.16 at 20 bps |
| Nearby parameter grid | All nine configurations were profitable |
| Deterministic 80% stock subsets | Ten subsets; minimum CAGR 20.55%, minimum Sharpe 1.095 |
| Single-sector exclusions | Every exclusion remained profitable |
| Random seed | Seeds 7, 42, and 137 met the point-estimate criteria |
| Training window | 1.0, 1.5, and 2.0 years met the point-estimate criteria |
| Multiple-testing adjustment | Deflated Sharpe probability estimate 90.64% across ten recorded trials |
| Block bootstrap | 2.5th percentile: 10.10% CAGR and -31.24% drawdown |
The bootstrap tail is materially weaker than the point estimate. This is not a contradiction: nearby-specification stability and path uncertainty measure different sources of risk.
The research design was fixed before the 2026-04-01 to 2026-07-21 holdout. No parameter was changed after observing this window.
| One-way cost | Total return | Annualized Sharpe | Maximum drawdown |
|---|---|---|---|
| 5 bps | 7.71% | 1.39 | -8.23% |
| 10 bps | 7.56% | 1.37 | -8.25% |
| 20 bps | 7.27% | 1.32 | -8.29% |
The annualized ratios are descriptive because the holdout contains only 76 sessions. They should not be treated as long-run performance estimates.
| Holdout component | Arithmetic contribution |
|---|---|
| Alpha Core gross P&L | -1.58% |
| Directional Risk Overlay gross P&L | +9.68% |
| Transaction costs | -0.14% |
| Arithmetic net sum | +7.96% |
| Compounded portfolio return | +7.71% |
The two sleeve contributions are arithmetic sums of daily P&L. The reported portfolio return compounds daily net returns, so +7.96% and +7.71% are not expected to match exactly.
The attribution identifies which portfolio decision drove the holdout. It does not claim that the overlay is a pure market-beta instrument or that returns have been decomposed into statistical alpha and beta.
The evidence supports a fixed, cost-aware portfolio process with explicit timing, exposure, and accounting controls. Development performance remains positive across the reviewed cost levels and nearby robustness checks. The one-time holdout also produced a positive combined return.
The component evidence is less conclusive. Alpha Core performance was positive during development but negative during the short holdout, while the Directional Risk Overlay generated the holdout gain. The next valid test is therefore not another adjustment to the observed holdout. It is to wait for genuinely new data and examine whether the fixed Alpha Core resumes positive contribution.
The repository includes a standard-library Python implementation of the cross-layer contracts on synthetic inputs:
public_reference/
├── data_contracts.py # cross-sectional schema and availability
├── contracts.py # information clock and walk-forward purge
├── research_pipeline.py # end-to-end sequencing and protected interfaces
├── portfolio_layers.py # neutrality, exposure, and sleeve composition
├── accounting.py # turnover, transaction costs, and P&L
├── audit.py # cross-layer decision audit record
└── walkthrough.py # synthetic end-to-end execution
tests/
├── test_data_contracts.py
├── test_public_reference.py
└── test_research_pipeline.py
The pipeline keeps three research components behind explicit interfaces:
observable feature panel
-> [private cross-sectional scorer]
-> validated full-universe scores
-> [private Alpha Core constructor]
-> validated neutral weights
-> [private directional policy]
-> validated overlay target
-> public execution, accounting, and audit contracts
This boundary preserves the private feature transformations, fitted models, portfolio optimizer, thresholds, and run artifacts. The surrounding code is executable rather than decorative pseudocode: private components cannot bypass the public checks on timing, coverage, neutrality, turnover, or reconciliation.
Run the synthetic walkthrough and 13 tests from the repository root:
python -m public_reference.walkthrough
python -m unittest discover -s tests -v- The universe is fixed and retrospective, so survivorship and membership-selection bias remain.
- Daily adjusted bars do not reconstruct intraday execution paths.
- Linear transaction costs do not capture nonlinear market impact, financing, borrow availability, taxes, or capacity constraints.
- Realized beta to SPY does not rule out other common-factor or regime exposure.
- A 76-session holdout is too short to establish long-run stability.
- The public reference code demonstrates research contracts on synthetic inputs; it did not generate the reported performance.
These boundaries limit the claims that can be made from the evidence. They do not change the reported development or holdout results.
| Path | Contents |
|---|---|
RESEARCH_NOTE.md |
Research question and core design decisions |
EVIDENCE_AND_LIMITATIONS.md |
Metrics, attribution, robustness, and interpretation limits |
docs/SYSTEM_DESIGN.md |
Stage-by-stage system specification |
docs/RESEARCH_LINEAGE.md |
Decision chain and evidence lineage |
public_reference/ |
Executable synthetic reference architecture |
tests/ |
Data, timing, portfolio, pipeline, and accounting tests |
DISCLOSURE.md |
Public/private implementation boundary |