Start with an offline demo using made-up data. Broker execution needs separate setup and supervised checks.
Retrospective research · fixed current pools · no live trading record
Review the original results, data availability and acquisition checks here. These records explain the evidence; they are not additional performance claims.
Fixed current pools: 10 US stocks, 10 Hong Kong stocks and 8 crypto assets. Stock data ends on 20 October 2026; crypto data ends on 30 September 2026. Historical pool membership, dividends and price adjustments, and executable fills remain limitations.
All fixed strategies at the predeclared nonzero one-way cost (US 20, HK 30, crypto 10 bps) for each market. Figures describe separate historical windows, not a compounded account or a forward test.
| Strategy | Windows | Positive excess | Median net | Worst net |
|---|---|---|---|---|
| Market benchmark | 11 | 0% (0/11) | 4.04% | -14.96% |
| Buy and hold | 11 | 72.73% (8/11) | 6.53% | -14.99% |
| Equal weight | 11 | 81.82% (9/11) | 7.15% | -14.85% |
| Positive momentum · 20 | 11 | 18.18% (2/11) | 0.35% | -21.26% |
| SMA · 20/60 | 11 | 54.55% (6/11) | 3.63% | -12.61% |
| Strategy | Windows | Positive excess | Median net | Worst net |
|---|---|---|---|---|
| Market benchmark | 10 | 0% (0/10) | 3.42% | -6.7% |
| Buy and hold | 10 | 70% (7/10) | 4.58% | -7.92% |
| Equal weight | 10 | 70% (7/10) | 3.78% | -8.09% |
| Positive momentum · 20 | 10 | 20% (2/10) | -0.81% | -22.68% |
| SMA · 20/60 | 10 | 50% (5/10) | 2.81% | -8.04% |
| Strategy | Windows | Positive excess | Median net | Worst net |
|---|---|---|---|---|
| Equal weight | 12 | 0% (0/12) | -0.18% | -31.96% |
| Positive momentum · 63 | 12 | 33.33% (4/12) | -0.12% | -46.09% |
| Trend · 200 | 12 | 50% (6/12) | 0.66% | -39.63% |
Positive excess counts only windows with a benchmark. Zero is not a win. Small, selected pools and retrospective windows limit generalization.
03 / VALIDATION SNAPSHOT
Fixed-input diagnostics checked on 19 October 2026. These samples do not establish deployable performance. Different universes, windows and accounting models cannot be ranked together.
HK equity-lab runs use one stock; US runs use SPY only. They cannot validate stock selection or cross-sectional ranking. Fixed costs are 30 bps for HK and 20 bps for US, per traded notional.
Image/CNN results can include in-sample predictions. No eligible walk-forward fold existed under the unchanged 1,260-day training window. A zero-position model is not a successful strategy.
| Strategy | Sample | Return | Sharpe | Drawdown | Evidence |
|---|---|---|---|---|---|
cnn_image_pattern | 2023-10-18 2026-09-30 | 22.77% | 1.539 | 4.65% | Includes in-sample risk |
cnn_pattern | 2023-10-18 2026-09-30 | 0.00% | N/A | 0.00% | No positions |
cnn_pattern_mvp | 2023-10-18 2026-09-30 | 22.77% | 1.539 | 4.65% | Includes in-sample risk |
cnn_walk_forward_rank | 2023-10-18 2026-09-30 | N/A | N/A | N/A | No eligible folds |
equal_weight | 2023-10-18 2026-09-30 | 152.99% | 1.479 | 22.12% | Limited universe |
momentum_20d | 2023-10-18 2026-09-30 | 6.91% | 0.717 | 3.84% | Limited universe |
random_turnover | 2023-10-18 2026-09-30 | 152.99% | 1.479 | 22.12% | Limited universe |
sma_cross | 2023-10-18 2026-09-30 | 13.62% | 1.068 | 4.65% | Limited universe |
hk_pure_momentum | 2024-11-01 2026-09-30 | 10.60% | 1.758 | 2.81% | Limited universe |
hk_momentum_volatility | 2024-11-01 2026-09-30 | 10.60% | 1.758 | 2.81% | Limited universe |
| Strategy | Sample | Return | Sharpe | Drawdown | Evidence |
|---|---|---|---|---|---|
cnn_image_pattern | 2023-10-03 2026-09-30 | 13.30% | 1.395 | 4.04% | Includes in-sample risk |
cnn_pattern | 2023-10-03 2026-09-30 | 0.00% | N/A | 0.00% | No positions |
cnn_pattern_mvp | 2023-10-03 2026-09-30 | 13.30% | 1.395 | 4.04% | Includes in-sample risk |
cnn_walk_forward_rank | 2023-10-03 2026-09-30 | N/A | N/A | N/A | No eligible folds |
equal_weight | 2023-10-03 2026-09-30 | 80.53% | 1.373 | 19.00% | Limited universe |
momentum_20d | 2023-10-03 2026-09-30 | 5.63% | 0.966 | 1.91% | Limited universe |
random_turnover | 2023-10-03 2026-09-30 | 80.53% | 1.373 | 19.00% | Limited universe |
sma_cross | 2023-10-03 2026-09-30 | 6.00% | 0.934 | 2.07% | Limited universe |
| Strategy | Sample | Return | Sharpe | Drawdown | Evidence |
|---|---|---|---|---|---|
global_pure_momentum | 2024-12-02 2026-09-30 | 31.95% | 1.436 | 12.71% | Native-currency proxy |
global_defended_rotation | 2024-12-02 2026-09-30 | 31.60% | 1.47 | 12.71% | Native-currency proxy |
Monthly HK diagnostics use the same one-stock subset at 10 bps. Regional rotation uses five ETFs at 10 bps, native-currency prices and a common calendar. FX, dividends, self-financing drift and point-in-time universes are not fully represented.
Archived 19–21 March 2026 events passed the recorded checks. Venue/product provenance and timezone are unverified. Maker fee is zero and latency is 100 ms. This is neither CoinPaprika L2 data nor verified live execution. Absolute PnL/drawdown have no funded-capital return denominator.
| Strategy | PnL in source quote units | Fills | Peak absolute inventory | Evidence |
|---|---|---|---|---|
B0 | -235.37 | 5,400 | 1 | Unverified-source replay |
B1 | -240.258 | 4,939 | 1 | Unverified-source replay |
B2 | -6.684 | 463 | 0.738 | Unverified-source replay |
B3 | -256.235 | 5,386 | 1 | Unverified-source replay |
A successful discovery request does not establish historical entitlement, point-in-time availability or a trading interface. Raw provider data and account details are excluded.
| Source | Current evidence | Stored observations |
|---|---|---|
aivix | Discovery works; historical PIT unknown | 2,280 |
index_one | Schema works; equity coverage unverified | N/A |
algoseek | No authorized configuration verified | N/A |
coinpaprika | Latest OHLCV works; daily price history stored | 730 |
ib_gateway | Partial history, then disconnected | 738 |
The read-only Gateway handshake and one HK history request succeeded, then error 1100 interrupted collection. One reconnect was refused. No order method was called.
Input receipts, experiment acceptance and remaining integration work ↗
Real daily bars and new diagnostic reruns. These are retrospective research results, not accepted research or execution targets.
Coverage dates span the earliest and latest observed bars across products, not a common start date. Product histories differ. The evaluated study remains fixed to 2023–2025.
binance_spot · Acquired · 8 Products · 8768 daily bars · 2023-01-01–2025-12-31 · Fixed study pool
binance_spot · Partial · 8 Products · 23778 daily bars · 2017-08-17–2026-09-30 · Fixed study pool
Both pools use 2023–2025 observations with 63 days of warmup. Equal weight buys once and holds; momentum selects positive trailing 63-day returns using only the previous close and otherwise holds cash. Costs are basis points per one-way traded notional. USDT is not USD. The disclosed fixed candidate pool has selection bias.
| Strategy | Products | Cost (bps) | Evaluation period | Net return | Max drawdown | Sharpe |
|---|---|---|---|---|---|---|
equal_weight | 8 | 0 | 2023-03-05 2025-12-31 | 200.29% | 50.08% | 0.93 |
equal_weight | 8 | 10 | 2023-03-05 2025-12-31 | 199.99% | 50.08% | 0.93 |
momentum_63d | 8 | 0 | 2023-03-05 2025-12-31 | 161.64% | 54.54% | 0.86 |
momentum_63d | 8 | 10 | 2023-03-05 2025-12-31 | 105.03% | 57.00% | 0.72 |
equal_weight | 2 | 0 | 2023-03-05 2025-12-31 | 190.95% | 40.25% | 1.01 |
equal_weight | 2 | 10 | 2023-03-05 2025-12-31 | 190.66% | 40.25% | 1.01 |
momentum_63d | 2 | 0 | 2023-03-05 2025-12-31 | 161.63% | 38.16% | 1.03 |
momentum_63d | 2 | 10 | 2023-03-05 2025-12-31 | 137.43% | 39.30% | 0.95 |
30 completed folds: 252 sessions of warmup and 63 evaluation sessions, with positions carried from each warmup. HK uses one stock; US uses SPY. These do not establish S&P500 stock-selection performance. Existing engine costs are 30 bps for HK and 20 bps for US; adjustment and availability timing remain unverified.
| Market | Strategy | Fold | Evaluation period | Net return | Max drawdown |
|---|---|---|---|---|---|
| HK | equal_weight | 1 | 2025-05-07 2025-08-04 | 10.60% | 6.80% |
| HK | momentum_20d | 1 | 2025-05-07 2025-08-04 | 1.65% | 1.38% |
| HK | sma_cross | 1 | 2025-05-07 2025-08-04 | 1.12% | 1.38% |
| HK | equal_weight | 2 | 2025-08-05 2025-11-04 | 12.03% | 9.02% |
| HK | momentum_20d | 2 | 2025-08-05 2025-11-04 | 0.64% | 1.40% |
| HK | sma_cross | 2 | 2025-08-05 2025-11-04 | 2.39% | 1.84% |
| HK | equal_weight | 3 | 2025-11-05 2026-02-04 | 28.70% | 8.05% |
| HK | momentum_20d | 3 | 2025-11-05 2026-02-04 | 3.05% | 1.65% |
| HK | sma_cross | 3 | 2025-11-05 2026-02-04 | 5.28% | 1.65% |
| HK | equal_weight | 4 | 2026-02-05 2026-05-13 | 1.51% | 19.28% |
| HK | momentum_20d | 4 | 2026-02-05 2026-05-13 | -0.98% | 2.17% |
| HK | sma_cross | 4 | 2026-02-05 2026-05-13 | -1.96% | 4.06% |
| HK | equal_weight | 5 | 2026-05-14 2026-08-13 | 15.52% | 8.08% |
| HK | momentum_20d | 5 | 2026-05-14 2026-08-13 | 0.91% | 1.89% |
| HK | sma_cross | 5 | 2026-05-14 2026-08-13 | 3.02% | 1.64% |
| US | equal_weight | 1 | 2025-04-07 2025-07-08 | 22.77% | 6.33% |
| US | momentum_20d | 1 | 2025-04-07 2025-07-08 | 1.81% | 0.53% |
| US | sma_cross | 1 | 2025-04-07 2025-07-08 | 0.84% | 0.55% |
| US | equal_weight | 2 | 2025-07-09 2025-10-06 | 8.26% | 2.41% |
| US | momentum_20d | 2 | 2025-07-09 2025-10-06 | 1.23% | 0.67% |
| US | sma_cross | 2 | 2025-07-09 2025-10-06 | 1.62% | 0.49% |
| US | equal_weight | 3 | 2025-10-07 2026-01-06 | 3.01% | 5.07% |
| US | momentum_20d | 3 | 2025-10-07 2026-01-06 | -1.06% | 1.49% |
| US | sma_cross | 3 | 2025-10-07 2026-01-06 | 0.63% | 1.02% |
| US | equal_weight | 4 | 2026-01-07 2026-04-08 | -2.28% | 9.13% |
| US | momentum_20d | 4 | 2026-01-07 2026-04-08 | -0.60% | 0.71% |
| US | sma_cross | 4 | 2026-01-07 2026-04-08 | -0.36% | 0.52% |
| US | equal_weight | 5 | 2026-04-09 2026-07-09 | 11.20% | 4.49% |
| US | momentum_20d | 5 | 2026-04-09 2026-07-09 | 0.98% | 1.38% |
| US | sma_cross | 5 | 2026-04-09 2026-07-09 | 1.03% | 0.91% |
Three archived days, 2026-03-19–21. Venue, product and account currency remain unknown. PnL is in recorded account units, not a funded percentage return. Baseline, 1 bps maker fee and 300 ms placement/cancellation latency change one factor at a time. A conservative-fill scenario is unsupported by the current model.
| Strategy | Scenario | PnL (unknown unit) | Fills |
|---|---|---|---|
| B0 | Baseline | -235.37 | 5400 |
| B0 | Maker fee 1 bps | -347.49 | 5400 |
| B0 | Latency 300 ms | -287.74 | 6293 |
| B1 | Baseline | -240.26 | 4939 |
| B1 | Maker fee 1 bps | -343.72 | 4939 |
| B1 | Latency 300 ms | -274.32 | 5890 |
| B2 | Baseline | -6.68 | 463 |
| B2 | Maker fee 1 bps | -16.29 | 463 |
| B2 | Latency 300 ms | -24.47 | 568 |
| B3 | Baseline | -256.23 | 5386 |
| B3 | Maker fee 1 bps | -368.12 | 5386 |
| B3 | Latency 300 ms | -291.69 | 6241 |
EXPLORE
This is a presentation of the offline tools. Switching paths changes the walkthrough; it does not run Python or place an order in your browser.
How a CNN reads price charts ↗
SYNTHETIC FIXTURE
Compare simple allocation rules and experimental CNN signals using a shared daily backtest. Start with equal weights, then inspect the assumptions behind each model.
The bundled example uses fictional HK symbols. U.S. ETF strategy ideas are a separate research candidate, not a validated U.S. trading demo.
These values are fictional inputs from a deterministic demo. Market prices and return results require separate data and calculation.
Model assumptions
Daily fractional-weight approximation. No board-lot handling, auction simulation, minimum commissions or FX conversion. Check run_status: a signal failure is marked as strategy_fallback, not a successful zero-return strategy.
From the repository root, set an absolute DATA_ROOT and an external UV_PROJECT_ENVIRONMENT first. The guide explains configuration. Demo generation refuses to overwrite existing data.
cd research/equity-lab
uv sync --locked --group dev
uv run --locked python -m equity_lab.cli demo
uv run --locked python -m equity_lab.cli backtest --strategy equal_weightSYNTHETIC FIXTURE
Trace visible queues, public trades and order latency. Compare four quoting variants while keeping the simulator and measurement assumptions explicit.
Bid 99.9, ask 100.1. The fixture contains 20 levels and 121 snapshots per date, across three fixed date labels.
These values are fictional inputs from a deterministic demo. Market prices and return results require separate data and calculation.
Model assumptions
Visible-queue simulation, not an exchange matching engine. Funding is a signal rather than settlement. Signal horizons require an exactly representable sampling grid; audit quality reports invalid numeric values. Commands importing configuration still require data settings before help. Overlapping-order queue realism remains unverified.
From the repository root, set an absolute DATA_ROOT and an external UV_PROJECT_ENVIRONMENT first. The guide explains configuration. Demo generation refuses to overwrite existing data.
cd research/crypto-mm
uv sync --locked --group dev
uv run --locked python -m crypto_mm.demo
uv run --locked python -m crypto_mm.commands.backtest_cli --strategy B0 --days 2026-03-19STRATEGY IDEAS
The research framework and the competition strategy are different stages. No competition strategy has been selected or registered.
A candidate uses crypto sentiment to inform a U.S. ETF portfolio. Data permissions, point-in-time coverage and reproducible out-of-sample comparison remain open. The equity lab supplies offline baseline experiments; it is not that completed strategy.
Read the candidate analysis ↗B0 starts from mid-price and inventory skew. B1 uses L1 microprice, B2 adapts spread to volatility, and B3 changes the inventory target using funding signals. Compare the models before drawing conclusions about their edge.
Read the simulation assumptions ↗ENGINEERING
The execution project separately validates versioned targets and lineage, applies Paper-first safeguards, and journals broker events. Demo reports are not executable target files.
A complete supervised order lifecycle remains unverified. The published v1 target schema lists JP/US while runtime normalization also recognizes HK; this existing mismatch needs resolution before relying on an HK handoff.
Explore execution readiness ↗