pareto_diophan

phase1_analysis.md

Phase 1: Manager-Level Performance Analysis

Pareto Technologies — Diophan

Analysis Date: 2026-03-17 Track Record: 2021-01-01 → 2024-12-31 (4.0 years, 1,225 daily observations) Asset Class: Digital Assets / Crypto Quantitative Trading Frequency: 7 days/week (~306 obs/yr) Benchmark: BTC-USD


Executive Summary

Diophan delivers extraordinary headline returns — a cumulative 12,842% over 4 years with a full-period Sharpe of 3.42 and max drawdown of only -21.9%. Factor regression across 6 asset classes explains virtually nothing (R² = 0.6–1.2%), placing 99%+ of return variance in the idiosyncratic bucket. Alpha is statistically significant at the HLZ t ≥ 3.0 threshold (t = 6.45), and the Deflated Sharpe survives even 50-trial adjustment.

However, the critical finding is a statistically significant alpha decay trend: annualized rolling alpha has declined from ~300% (2021) to ~50–100% (2023–24), with a regression slope p-value < 0.0001. This is the single most important risk to the allocation thesis.

Verdict Summary

TestResultSignal
Alpha genuine?✅ R² = 0.6%, %idio = 99.4%Pure alpha, not factor bet
Track record significant?✅ t(α) = 6.45 > 3.0 (HLZ)Highly significant
Backtest credible?✅ DSR = 1.0 even at 50 trialsNot overfitting
Style drift?✅ No material factor driftConsistent near-zero betas
Alpha decaying?⚠️ YES — declining trend p < 0.001Fading edge signal
Hidden factor exposure?✅ None detected across 6 factorsNo hidden bets
Nonlinear risk?⚠️ Neg. skew (-0.82) + fat tails (kurt=3.2)Left-tail risk present
Replicable by indices?✅ RBSA R² < 0 — not replicableGenuinely unique

Step 1: Basic Risk-Adjusted Returns

MetricValue
Annualized Return237.0%
Annualized Volatility36.6%
Sharpe Ratio3.42
Sortino Ratio2.84
Calmar Ratio10.83
Max Drawdown-21.9%
Skewness-0.815
Excess Kurtosis3.193
Daily Mean Return0.42%
Daily Std Dev2.09%

Year-by-Year Performance

YearReturnVolSharpeMax DDObs
2021928.7%40.9%5.84-9.2%323
2022224.1%41.2%2.98-21.9%310
202356.6%27.6%1.66-18.7%280
2024147.8%34.3%2.72-19.1%312

Interpretation: The Sharpe of 3.42 is in the top decile of any strategy class. However, negative skewness (-0.82) combined with excess kurtosis (3.19) indicates a payoff profile with more frequent small gains but exposure to occasional large losses — consistent with short-vol or mean-reversion strategies in crypto. This warrants monitoring of tail risk. [R2, R3]

The year-by-year trajectory is notable: 2021 was exceptional (Sharpe 5.84), followed by a significant step-down in 2023 (Sharpe 1.66), with partial recovery in 2024. This pattern is consistent with alpha decay as the strategy's edge is potentially arbitraged.


Step 2: Drawdown Analysis

RankMax DDStartTroughRecoveryDurationDeclineRecovery
1-21.9%2022-09-082022-10-032022-11-1568d25d43d
2-19.1%2024-01-092024-03-042024-03-2374d55d19d
3-18.7%2023-01-272023-03-062023-04-1982d38d44d
4-14.2%2022-02-082022-02-112022-02-2820d3d17d
5-14.2%2023-06-012023-08-282023-10-21142d88d54d

Interpretation: 109 total drawdown episodes over 4 years. The worst (-21.9%) recovered in 43 days. All top-5 drawdowns recovered within 3 months, suggesting disciplined risk management or mean-reverting alpha. However, the 5th-ranked drawdown (Jun–Oct 2023) had an 88-day decline phase — the longest — which coincides with the weakest performance year, consistent with the alpha decay thesis.


Step 3: Rolling Metrics

WindowMean SharpeMedianMinMax% Positive
3-month3.343.28-2.769.3489.0%
6-month3.183.090.168.04100.0%
12-month2.862.740.796.23100.0%

Interpretation: Rolling 6m and 12m Sharpe have never gone negative — an exceptional consistency record. The 3m window shows brief periods of negative Sharpe, but these represent transient drawdown episodes rather than structural breaks. The declining trajectory of mean rolling Sharpe (3.34 → 3.18 → 2.86 as window lengthens) reflects the alpha decay from 2021's peak.

Rolling Sharpe


Step 4: Capture Ratios vs BTC

MetricValue
Up Capture20.0%
Down Capture-15.5%
Capture Ratio-1.29

Interpretation: The negative down capture is highly significant — when BTC declines, Diophan on average profits. This is a convex payoff profile vs the crypto market. The strategy captures only 20% of BTC upside but generates positive returns during BTC drawdowns, suggesting the strategy is genuinely market-neutral to slightly contrarian. The traditional capture ratio metric doesn't apply cleanly here (negative denominator), but the asymmetric profile is unambiguously favorable for portfolio construction.


Step 5: VaR / CVaR

MetricValue
VaR (95%)-3.66% daily
CVaR (95%)-5.10% daily
VaR (99%)-5.61% daily
CVaR (99%)-7.11% daily

Interpretation: The CVaR/VaR ratio at 99% is 1.27, indicating moderate tail thickness beyond VaR. A -5.1% daily CVaR at 95% on a strategy with 0.42% daily mean return implies a "break-even" metric of roughly 12 days — the strategy earns back its expected worst-day loss in under two weeks of normal performance.


Step 6: Factor Regression — CRITICAL ✅

Model 1: Single-Factor (BTC)

MetricValue
Alpha (ann.)120.7%
Alpha t-stat6.43
β(BTC)0.028 (t=1.07, NS)
0.0021
% Idiosyncratic99.8%

Model 2: Crypto 3-Factor (BTC + ETH-BTC Spread + BTC Momentum)

MetricValue
Alpha (ann.)124.8%
Alpha t-stat6.45
β(BTC)0.031 (t=1.18, NS)
β(ETH-BTC)-0.052 (t=-1.65, NS)
β(BTC_MOM)-0.004 (t=-0.98, NS)
0.0064
% Idiosyncratic99.4%

Model 3: Expanded 6-Factor (BTC + ETH + SPY + GLD + TLT + USD)

MetricValue
Alpha (ann.)135.1%
Alpha t-stat5.79
0.012
% Idiosyncratic98.8%

Only ETH shows marginal significance (t=-2.09, p=0.036) but falls below the HLZ threshold of 3.0. The joint F-test is not significant (p=0.264). No factor — crypto or traditional — explains the strategy's returns.

Interpretation [B1, R1]: With R² < 0.02 across all specifications, this is definitively NOT a factor bet. The strategy's return stream is orthogonal to BTC, ETH, equities, gold, bonds, and USD. Alpha of 120–135% annualized with t-stats of 5.8–6.5 survives even the most stringent significance thresholds. HAC standard errors with 6 lags account for any autocorrelation in the residuals.


Step 7: % Idiosyncratic Variance — CRITICAL ✅

MetricValueThreshold
% Idiosyncratic99.4%≥ 75%
% Factor0.6%
Implied IR3.43

Interpretation [B1 Insight 7.1]: At 99.4% idiosyncratic variance, the strategy's Sharpe ratio is essentially equal to its Information Ratio (SR = IR × √0.994 = 3.42). There is no factor dilution whatsoever. This is the theoretical ideal for alpha generation — maximum leverage capacity with zero factor drag.


Step 8: Expanded Factor Model (Crypto + TradFi)

Individual Factor Correlations

FactorCorrelation
BTC+0.022
ETH-0.025
SPY+0.022
GLD+0.054
TLT+0.050
USD-0.049

All correlations are below ±0.06. The strategy is essentially uncorrelated with every major asset class — an ideal diversifier. The marginal ETH exposure (β=-0.074, t=-2.09) likely reflects the strategy's crypto universe selection rather than a deliberate directional bet.


Step 9: Rolling Factor Analysis ⚠️

Rolling 180-Day Alpha (Annualized)

MetricValue
Mean112.8%
Median108.2%
Min-16.7%
Max309.5%
Std72.6%

Alpha Trend Analysis

MetricValue
Linear slope-0.0000054 per day
Annualized decline~50%/year
R² of trend0.468
P-value< 0.0001

⚠️ ALPHA DECAY DETECTED: The rolling alpha has declined from ~300% (mid-2021) to ~50–100% (2023–24) with a statistically significant negative trend (p < 0.0001, R² = 0.47). This is consistent with the Adaptive Markets Hypothesis [B1] — the strategy's edge is being arbitraged over time as competitors enter the crypto quant space.

Rolling BTC Beta

Mean beta is 0.037 (essentially zero), with a standard deviation of 0.058. The strategy oscillates around market-neutral, with brief excursions to ±0.15 that don't persist — no style drift detected in the traditional sense.

Rolling Factor Analysis


Step 10: RBSA — Sharpe Style Analysis ✅

Full-Period Style Weights (Constrained)

Style IndexWeight
BTC8.4%
ETH0.0%
SPY28.7%
GLD31.2%
TLT31.6%

RBSA R² = -0.075 (effectively zero)

Interpretation [R4:style_analysis_no_holdings]: The negative R² means the constrained combination of standard asset classes does worse than simply predicting the mean. The strategy is not replicable by any long-only combination of crypto, equity, gold, or bond indices. The style weights shown above are artifacts of the constraint forcing weights to sum to 1 — they have no economic meaning when R² is negative.

RBSA Style Weights


Step 11: Statistical Significance Gates ✅

Harvey-Liu-Zhu (2016) Threshold Tests

| Parameter | Value | |t|| Threshold | Verdict | |-----------|-------|------|-----------|---------| | Alpha | 124.8% ann. | 6.45 | ≥ 3.0 (HLZ) | ✅ SIGNIFICANT | | β(BTC) | +0.031 | 1.18 | ≥ 3.0 | ❌ Not significant | | β(ETH-BTC) | -0.052 | 1.65 | ≥ 3.0 | ❌ Not significant | | β(BTC_MOM) | -0.004 | 0.98 | ≥ 3.0 | ❌ Not significant |

Diagnostic Tests

TestStatisticP-valueVerdict
N/T Ratio0.0025✅ Well within bounds
Ljung-Box (10 lags)13.090.219✅ No autocorrelation
Breusch-Pagan3.800.284✅ No heteroskedasticity

Interpretation: Alpha passes even the strictest multiple-testing threshold. No factor exposure is significant. Residual diagnostics confirm the regression is well-specified — no autocorrelation or heteroskedasticity inflate significance.


Step 12: Deflated Sharpe Ratio ✅

Deflated Sharpe (Bailey & Lopez de Prado)

Trials AssumedDSRp-valueVerdict
11.0000.000✅ PASS
51.0000.000✅ PASS
101.0000.000✅ PASS
201.0000.000✅ PASS
501.0000.000✅ PASS

Haircut Sharpe (Harvey & Liu)

TrialsHaircutAdjusted SR
10.0%3.42
521.4%2.69
1026.9%2.50
2031.7%2.34
5037.4%2.14

Minimum Track Record Length

MetricValue
MinTRL (95%)120 obs (0.4 years)
Actual track record1,225 obs (4.0 years)
Verdict✅ 10× required length

Interpretation [R8]: Even with the most aggressive multiple-testing haircut (50 trials, 37.4% haircut), the adjusted Sharpe of 2.14 remains exceptional. The track record is 10× the minimum required length. This is not a backtest artifact.


Step 13: Additional Analysis

Monthly Returns (%)

JanFebMarAprMayJunJulAugSepOctNovDec
202139.025.130.118.634.652.23.17.926.78.86.914.2
202213.518.712.67.429.820.618.15.1-15.9-4.021.44.7
20236.0-6.9-1.821.921.8-7.3-3.80.58.34.56.90.3
2024-2.5-4.717.1-3.99.513.916.025.611.4-0.83.513.9

Negative months: Only 9 of 48 months (18.8%) are negative. The worst month was Sep 2022 (-15.9%).

Cumulative Performance Comprehensive Tear Sheet


Judgment Framework — Final Assessment

1. Is alpha genuine? ✅ YES

  • R² = 0.6% (single-factor) to 1.2% (6-factor) → NOT a factor bet [B1]
  • % Idiosyncratic = 99.4% → far above 75% threshold [B1 Insight 7.1]
  • No hidden factor exposures across crypto or traditional asset classes

2. Track record significant? ✅ YES

  • Alpha t-stat = 6.45 → exceeds both standard (2.0) and HLZ (3.0) thresholds [R4, B5]
  • Track record is 10× the minimum required length
  • No autocorrelation or heteroskedasticity inflate significance

3. Backtest credible? ✅ YES (Live Track Record)

  • DSR > 0.95 at all trial counts up to 50 [R8]
  • Haircut Sharpe remains above 2.0 even at 50 trials
  • This is a live track record, not a backtest — but the DSR/HSR analysis confirms it would survive scrutiny even if it were

4. Style drift? ✅ NO

  • Rolling BTC beta oscillates ±0.06 around zero — no persistent directional tilt
  • No material style drift detected [R1, B3]
  • Strategy maintains consistent near-zero market exposure

5. Alpha decaying? ⚠️ YES — CRITICAL

  • Rolling alpha trend has negative slope with p < 0.0001 and R² = 0.47 [R1]
  • 2021 Sharpe of 5.84 → 2023 Sharpe of 1.66 → 2024 Sharpe of 2.72
  • Consistent with Adaptive Markets Hypothesis: crypto quant alpha is being competed away [B1]
  • Mitigant: 2024 showed recovery (2.72 vs 2023's 1.66), suggesting alpha may have stabilized or the strategy adapted

6. Hidden factor exposure? ✅ NONE

  • Correlation with all 6 tested factors < ±0.06
  • No factor beta significant at HLZ threshold
  • Strategy is genuinely orthogonal to markets

7. Nonlinear risk? ⚠️ MODERATE

  • Negative skew (-0.82) + excess kurtosis (3.19) → fat left tail [R2, R3]
  • CVaR(95%) = -5.1% daily — 12× daily mean return
  • Not as extreme as short-vol strategies, but warrants position sizing discipline

8. Replicability? ✅ NOT REPLICABLE

  • RBSA R² < 0 — no combination of standard indices can replicate
  • Genuinely unique return stream [R4:style_analysis_no_holdings]

Risk Flags & Monitoring Recommendations

🔴 High Priority

  1. Alpha Decay Trend — Monitor rolling 12m Sharpe. If it drops below 1.5 for 2+ consecutive quarters, reassess allocation thesis. The 2021→2023 decline was steep; 2024's recovery needs to be sustained.

🟡 Medium Priority

  1. Tail Risk — Negative skew with fat tails means VaR underestimates true risk. Position size should account for CVaR, not VaR. Consider maintaining a GSAM Green Zone monitoring regime.
  2. Return Magnitude — Returns of this magnitude (237% annualized) raise questions about capacity constraints, sustainability, and operational risk. Due diligence should probe: What is AUM? What is estimated capacity? Is leverage involved?
  3. Crypto-Specific Risks — Exchange counterparty risk, smart contract risk, regulatory risk. These are not captured in return-based analysis.

🟢 Low Priority

  1. Factor exposure — Near-zero and stable. Continue monitoring for drift.
  2. Drawdown recovery — All drawdowns have recovered quickly. Pattern is healthy.

Data Quality Notes

  • Track record trades 7 days/week, consistent with crypto markets
  • 1,225 observations over 4 years (~306/year)
  • No gaps detected; some days may be interpolated (dates like Jan 1, Jan 3 include weekends)
  • Returns are net-of-fee (as ingested)
  • Benchmark alignment: BTC-USD matched on exact dates, 1,225 common observations

Files Generated

FileDescription
step3_rolling_sharpe.pngRolling 3m/6m/12m Sharpe ratio charts
step9_rolling_factor.pngRolling alpha and beta with drift detection
step10_rbsa.pngRolling RBSA style weights
step13_tear_sheet.pngComprehensive 7-panel tear sheet
cumulative_performance.pngCumulative return with drawdowns
quantstats_report.htmlFull QuantStats interactive report
step1_5_results.jsonComputed metrics (Steps 1–5)
step6_7_factor_results.jsonFactor regression results

Analysis conducted using: quantstats, statsmodels (OLS with HAC), scipy (RBSA optimization), matplotlib. All computations from raw daily returns; no numbers are assumed or interpolated.