Crypto Research · satgas.ai.id
Validation of ma_filter_breakout (E18)
One strategy passed the initial gate. This page retests it against a stricter
standard: whether any future information leaked, whether the results hold up after
costs are raised, whether the parameters were chosen too cleverly, and whether the numbers still make
sense after correction because we have tried many combinations.
1. Look-ahead test (prefix)
Signals are recalculated on data truncated at 40 random points. If the signal on truncated
data differs from the full-data signal, it means future information has leaked.
| Points tested | 40 |
| Different signals | 0 |
| Verdict | no leakage |
2. Cost sensitivity
| Cost per transaction | Sharpe |
Return | CAGR | Max DD |
|---|
| 0 bps | 0.76 | 152.5% | 16.6% | -28.1% |
| 5 bps | 0.73 | 144.8% | 16.0% | -28.4% |
| 10 bps | 0.71 | 137.3% | 15.4% | -28.7% |
| 15 bps | 0.69 | 130.1% | 14.8% | -29.0% |
| 25 bps | 0.65 | 116.2% | 13.7% | -29.5% |
| 40 bps | 0.59 | 97.0% | 11.9% | -30.4% |
A cost of 15 bps is the base assumption (10 fee + 5 slippage). If Sharpe collapses
at 25 bps, this strategy is fragile to real execution conditions.
3. All parameters, not just the winner
| Parameter | Sharpe | Return |
CAGR | Max DD | Exposure |
|---|
| {"entry": 20} | 0.69 | 130.1% | 14.8% | -29.0% | 24% |
| {"entry": 55} | 0.64 | 123.8% | 14.3% | -27.9% | 28% |
| {"entry": 200} | 0.39 | 53.3% | 7.3% | -43.4% | 36% |
| {"entry": 100} | 0.37 | 46.7% | 6.6% | -28.8% | 32% |
If only one value is good and the rest are bad, that is a sign of overfit.
A neat spread across several values is more convincing.
4. Last 12-month holdout test
| Parameters frozen from data before 2025-10-01 | {"entry": 20} |
| Test period | 388 days |
| Strategy return | 0.2% |
| BTC return over the same period | -32.4% |
| Strategy vs BTC Sharpe | 0.07 vs -0.62 |
| Strategy drawdown | -13.3% |
Parameters were selected only from data before the cutoff date, so the numbers here
have never been used to choose anything. This is the test closest to real conditions.
5. Data frequency (weekly bars)
| Weekly Sharpe | 0.11 |
| Return | 3.1% |
| Drawdown | -34.0% |
| Number of bars | 315 |
If the effect only appears on daily bars but disappears on weekly bars,
it is most likely micro noise, not an economic edge.
6. Universe dependence
| Universe | Sharpe | Return |
Max DD |
|---|
| BTC, ETH | 0.69 | 130.1% | -29.0% |
| BTC, ETH, SOL | 0.69 | 130.1% | -29.0% |
The universe uses coins that still exist today. Dead coins have no
exchange data, so the results contain survivorship bias and should be read as
an upper bound, not a number to be expected.
7. Null test with two block lengths
| Block | Original Sharpe |
Null median | p-value |
|---|
| 10 days | 0.69 |
0.47 | 0.235 |
| 30 days | 0.69 |
0.62 | 0.407 |
Two block lengths are used to ensure the conclusion does not depend on
one test parameter choice. p below 0.05 means it is hard to explain by random data.
8. Deflated Sharpe Ratio
| Daily Sharpe | 0.0362 |
| Sharpe expected from luck (SR0) | 0.0977 |
| DSR (probability of a real edge) | 0.002 |
| Return skew / kurtosis | 0.86 / 26.26 |
DSR corrects Sharpe because we have tested dozens of parameter combinations.
A value below 0.95 means the claim of an edge is not yet strong enough to be considered certain.
Conclusion
Look-ahead: clean. At a cost of 15 bps, Sharpe 0.69 and CAGR 14.8%. At a cost of 25 bps, Sharpe 0.65. 12-month holdout: strategy +0.2% vs BTC -32.4%. Null test significant on 0 of 2 block lengths. DSR 0.00 (needs >= 0.95 for a strong claim). Conclusion: the evidence is not yet sufficient to consider this a durable edge.
Updated 25 Oct 2026 07:21 UTC+7. All figures are calculated
from data stored in research/db. There is no real money, no orders.