C++20 research and execution framework for prior-only intraday strategies, volatility-based allocation, shared-capital accounting and synthetic broker recovery across SPY, QQQ, IWM, TLT and GLD.
This project studies whether simple intraday signals retain an economically meaningful advantage once their information sets, transaction costs, execution delays and market exposure are made explicit. It also implements the account and order machinery needed to translate those decisions into a recoverable execution process.
The research combines late-session momentum, volume-conditioned continuation and opening-anchor mean reversion, with a separate daily-risk allocation study. The registered archive spans 2016-2025. Development is followed by a chronological 2024 validation and a frozen 2025 final retrospective evaluation.
The C++ implementation owns signal estimation, financial accounts, order transitions, matching and journal recovery. Historical market-data acquisition, independent broker simulation and recorded-result publication have separate responsibilities. Python is limited in the active tree to build orchestration, issuer-data preparation and publication utilities.
The completed evaluation retains both positive and negative findings. In 2025, volume-conditioned continuation and reversion have small positive pooled index means, but both depend on their largest winning date. The primary shared account finishes below its starting capital, and the registered uncertainty and risk-matching results do not support strategy advancement.
The technical report contains the full mathematics, experimental conventions and results. Its standalone HTML version embeds saved synthetic event views and opens without a server or JavaScript. The original SPY release retains its separate 2022-2023 sample and findings.
- Visual Summary
- Research Questions
- Dataset and Study Design
- Pipeline Design
- Strategy Models
- Risk and Exposure Matching
- Shared Capital and Broker Lifecycle
- Key Results
- Statistical Interpretation
- Runtime Measurements
- Implementation and Reproduction
- Repository Structure
- References
Equal-weight research-index means at 1 bp per side and one-minute delay, with final primary paired-date coverage. Each period retains its actual observation mask. These indices are separate from continuously funded portfolio wealth.
Dots show estimates and left ticks show block-40 numerical lower bounds. Gold lines mark the registered material increments. The panels have different units; calibration size failures limit their confidence interpretation.
The candidate and four control accounts each start with USD 100,000 and complete 250 evaluation dates in 2025. Their different market exposures remain visible.
The original broker demonstration records order, cash and reservation states. The expanded study additionally tests an independently journaled native child broker. Both use constructed events rather than observed external-broker fills.
The project separates four questions that require different evidence:
- Directional information: do the intraday rules improve outcomes relative to controls using comparable decision and holding conventions?
- Risk allocation: does intraday-aware risk estimation improve forecasting loss or allocation value over a daily-only reference?
- Economic robustness: do differences survive costs, delayed references, concentration diagnostics and predefined exposure matching?
- Execution correctness: can concurrent strategy ownership, partial fills, cancellations and interrupted delivery be reconciled without changing balances or applying an execution twice?
A valid recovery trace does not establish a profitable signal. A positive research-index mean does not establish a successful funded portfolio.
| Component | Specification |
|---|---|
| Instruments | SPY, QQQ, IWM, TLT and GLD |
| Market observations | Historical minute bars, exchange calendars and corporate-action evidence |
| Acquisition | Retained Alpaca historical responses with pagination, checkpoints and immutable identities |
| Archive horizon | 2016-2025; 2,514 exchange sessions |
| Initialization | 2016 and subsequent available history before each decision |
| Development | 2017-2023; 1,760 calendar dates |
| Validation | 2024; 252 evaluation dates |
| Final evaluation | 2025; 250 evaluation dates |
| Historical cost scenarios | 0, 0.5, 1, 2 and 5 bp per side |
| Execution-reference delays | One and two minutes |
The universe is selected and includes correlated equity exposures. Earlier GLD research and inspected SPY outcomes mean the archive is not globally unseen. The chronological reservation controls selection using this extension's own later-period outcomes; it does not erase prior market or literature exposure.
The original complete-minute intersection fails its development-year coverage gate. A separately registered input amendment distinguishes required operational execution slots from feature availability. The parent failure remains recorded.
Three missing SPY/IWM/GLD closes on 12 August 2019 remain unavailable after bounded recovery. Daily histories reset across the gap. Required missing prices stop the affected capital path; later recovery of model history does not restart that account or fill its missing tail.
Canonical cash events retain issuer evidence, ex-dates and payment dates. Duplicate announcements are not summed. Split-absence and GLD cash-absence cross-checks retain their stated completeness assumptions. The study design records source amendments, masks, prior exposure and the original-SPY provider-vintage comparison.
Raw responses, calendars, action references and eligibility masks are bound to a registered study manifest. Missing prerequisites stop registration. An updated provider vintage cannot silently overwrite an accepted snapshot.
Signals use completed observations. Rolling estimates use strictly earlier sessions, while execution references occur after the decision timestamp. Unavailable features and unsupported model fits remain dated outcomes.
The same primary decisions feed the fixed cost/delay grid. Control policies and fourteen one-factor history, observation-grid and holding-period variations remain in the comparison. No alternative is selected from final results.
Fractional research accounts retain cash, shares and corporate-action claims. Whole-share execution accounts additionally maintain reservations, explicit strategy ownership, target netting and aggregate admission limits.
Year, instrument, prior-only market-condition and holding-period views accompany concentration, turnover, drawdown, exposure and costs. Joint calendar resampling retains paired comparisons and unavailable observations. Controlled synthetic experiments assess the inference procedure separately.
Verified 2024 results authorize a copied executable and specification freeze before 2025 access. Final capital starts fresh while model histories remain warmed from earlier observations. The accepted replacement runner reproduces all 26 financial artifact identities and four 2024 scientific summaries before its final evaluation. The acceptance record retains both freeze identities and the payment-date implementation correction.
The late-session rule observes the completed move from the opening reference and enters long only after a positive signal. Its comparison includes a same-window always-long control, so avoiding market exposure can be distinguished from earning an incremental return over that exposure.
Continuation compares the current opening-relative move with the mean absolute same-clock move over twenty earlier sessions:
The volume-conditioned variant requires cumulative volume above its prior same-clock median for fresh entries. Volume does not determine exits. The price-only continuation policy retains the directional reference without this additional entry condition.
Five-minute log residuals around a fixed opening anchor follow the estimated relation
For an admissible positive sub-unit slope, the corresponding OU parameters and expected delayed holding-period move are
The candidate compares this expected move with the fixed cost hurdle. Sign-only, drift and always-long controls separate the fitted reversion effect from simpler entry rules and market exposure. Numerical OU admissibility does not establish stationarity or identify a latent diffusion.
The constructed finite-sample and observation-noise study examines that identification problem directly:
For a round trip with entry and exit references
The historical comparison retains 400 instrument/policy/cost/delay series and 240 paired contrasts. These charges are explicit hypothetical scenarios, not measured all-in spreads or broker fees.
A separate fixed historical SIP quote sample contains 910 windows, of which 232 meet the registered condition and timestamp rules. Its crossing-cost references do not identify market impact, queue position or synthetic fill probabilities. Sparse or ambiguous observations remain unestimated.
The separate daily-risk study compares strictly lagged daily-only and
intraday-aware estimates. Forecasting loss and subsequent account economics are
evaluated separately. For forecast variance
Allocation comparisons include annualized certainty-equivalent return, with risk-aversion coefficient three:
The predefined exposure comparison uses 252 prior session slots. Candidate and control scale toward the smaller of their estimated annual volatilities and a fixed 10% target:
The implementation retains missing slots and unsupported sizing. Forty-five unit-reference accounts feed fifty funded matched accounts. This comparison asks whether an apparent advantage survives a predefined exposure convention; it does not infer risk-matched alpha from separately masked return means.
One whole-share USD account supports five inventories and fifteen strategy owners. Owner claims, target netting and aggregate limits operate on actual balances. Daily-risk allocation remains a separate experiment rather than a fourth directional owner.
The independently journaled synthetic broker runs in a native child process. Its in-process and separate-process routes share the matching core. The lifecycle covers:
- DAY, IOC and GTC limit orders with finite quote liquidity;
- acknowledgments, partial fills, fees and cancellation races;
- managed cancel/new replacement and stable execution identities;
- disconnected delivery, interrupted commits and activity-history recovery;
- reconciliation of independent broker and client accounts before routing resumes.
The process study records 313 checks in twelve scenario families across 21 child processes. A separate execution study retains 680 account cases and 510 matched contrasts, with every terminal broker/client account reconciled. These case counts describe synthetic experimental coverage, not market observations.
OrderRuntime, SharedCapitalBook and their journal owners maintain exact
integer monetary values, whole shares and reservations. Journal schemas declare
their units, including cents and USD 1e-9. Candidate state is synchronized before
dispatch eligibility is published. Recovery replays the same financial rules;
contradictory snapshots cannot overwrite the ledger.
Anonymous pipes, constructed quote budgets and simulated fills do not establish external-broker connectivity or real-market execution quality. Historical Alpaca data acquisition remains available without a paid feed or live broker service. The architecture and broker contract define ownership, persistence and recovery boundaries.
At 1 bp per side and one-minute delay, each pooled intraday index has 247 observed common dates out of 250 final evaluation sessions:
| Research index | Net mean, bp per observed date | Mean after zeroing its largest winner, bp/date |
|---|---|---|
| Momentum | -0.7684 | -1.0026 |
| Volume-conditioned continuation | +0.3157 | -0.7021 |
| Reversion candidate | +0.1042 | -0.0961 |
Winner removal retains the original denominator and does not refit a strategy or simulate another funded account. Both positive point means become negative under that diagnostic, demonstrating concentration in the recorded index paths.
All 200 risk-account cases and fifty shared-capital cases complete 2025. For the primary shared candidate:
| Metric | Recorded result |
|---|---|
| Starting capital | USD 100,000 |
| Terminal equity | USD 99,763.25 |
| Funded return | -0.2367% |
| Charged transaction costs | USD 1,759.59 |
| Mean market exposure | 23.9159% |
| Observed-path drawdown | 1.0196% |
| Observed dates | 250 |
All four shared control accounts also complete the year. Their distinct exposures remain part of interpretation. The positive continuation and reversion index means therefore do not imply a profitable combined funded account.
Earlier development ledgers and the primary 2024 shared account have stopped tails. The latter records only 79 observed dates. Complete 2025 paths do not repair those earlier gaps or retroactively pass the original SPY screens.
Paired full-calendar moving-block resampling retains missing slots and uses
joint draws across the six registered endpoints. With HAC standard error
Blocks 5, 10, 20 and 40, the complete endpoint family and material increments remain fixed. Final primary coverage meets the registered count and fraction requirements for all six endpoints, including 238/250 QLIKE pairs.
Two qualifications govern the conclusion:
- Material improvement remains unestablished. No primary numerical lower bound exceeds its material increment at any registered block length. A positive lower bound for QLIKE is still below its specified material threshold.
- Inference remains limited. The 20,000-family controlled calibration detects size excess in 19/20 case/block cells, preventing a validated nominal 95% simultaneous-confidence interpretation. The matched family is unavailable because reversion versus drift has zero paired observations; that endpoint is retained rather than dropped.
The study supports a completed research and engineering evaluation, with no strategy advancement under the registered design. The source-bound endpoint table and final acceptance record retain estimates, denominators, unavailable values and source identities.
The archived Release benchmark contains 120 rows across eight modes, with five repetitions at 8, 32 and 128 round trips. At 128 round trips, the Windows median trial p50 measurements are:
| Processing mode | Median trial p50 |
|---|---|
| Direct core calls | 24.4 us |
| Synchronized typed journals | 2.4384 ms |
| Synchronized paper journals | 3.5602 ms |
These timings measure the original bounded single-symbol workload, not the expanded five-instrument peer or end-to-end market latency. Raw samples, compiler/host identities and variability remain recorded. Cross-host Windows/Linux differences cannot isolate an operating-system effect. The benchmark specification defines timer scope, warm-cache recovery, instrumentation and resource limits.
Requirements are Python 3.10+ and C++20. Windows uses Visual Studio 2022/2026
with Desktop development with C++ and a Windows SDK. Linux uses GCC/Clang, ar
and OpenSSL development headers. The first build acquires pinned nlohmann/json
3.12.0; --json-include accepts a preinstalled copy.
py -3 -B build.py --configuration Release --target project_demo
& .\scratch\native\Release\project_demo.exe
py -3 -B research/verify_publication.pyIn Visual Studio, open Runtime.sln, select Release | x64 and build.
Ctrl+F5 runs the default demo. On Linux, use python instead of py -3,
forward-slash executable paths and no .exe suffix. Debug supports stepping
through code; Release provides the performance reference.
The offline demonstration needs no market credentials or private dataset. Full historical reproduction additionally requires the registered raw archive, issuer/action references, calibration and original/replacement frozen-run evidence. Public summaries cannot reconstruct omitted observations.
The active Python files serve build, issuer-data and publication tasks. The complete independent numerical verification suite is preserved at the verified publication revision. The runbook provides its pinned checkout and execution commands. The original report's scientific verification commands apply to that complete revision.
| Location | Responsibility |
|---|---|
src/research |
Prior-only models, registered comparisons, risk estimates and inference |
src/accounting, src/execution |
Fractional/whole-share ledgers, ownership, matching and recovery |
src/market, src/alpaca, src/storage, src/platform |
Validated observations, archive identities and platform resources |
src/reporting, apps |
Result-bound reports and thin application composition |
research |
Frozen protocols, issuer-data preparation and one publication builder |
docs |
Methods, source-bound tables, figures and the standalone report |
scratch |
Ignored datasets, native builds and full run evidence |
There is one authoritative C++ implementation of the financial engines. Domain owners enforce state and resource invariants, while pure formulas remain small functions. The independent Python oracles are verification references. The current source tree uses one build driver, without CMake/CTest or a hosted workflow.
Material limits include prior research exposure, selected and correlated instruments, historical final-bar information, hypothetical costs, no cash yield, uncalibrated synthetic liquidity and bounded journals. Frozen protocols and observed failures remain part of the completed result.
- Gao, Han, Li and Zhou (2018), Market intraday momentum.
- Moreira and Muir (2017), Volatility-managed portfolios.
- Kunsch (1989), The jackknife and the bootstrap for general stationary observations.
- Registered research inputs and full technical report.
Kartik Vij