Skip to content
QuantCalc Tools
Directory

Cagr Irr Drawdown Compounding

Continuous Log Returns vs Simple Returns: Time Additivity vs Asset Weights

Logarithmic returns are additive across time but NOT across portfolio assets; multiplying weights w_i by ln(P_t/P_{t-1}) misstates volatile multi-asset returns and breaks VaR/CAGR models.

By Elena Rostova, FRM, CQFReview not recorded · Updated
Continuous Log Returns vs Simple Returns: Time Additivity vs Asset Weights — QuantCalc Tools 2026 quantitative & supervisory data card
QuantCalc Tools (CAGR vs IRR, Volatility Drag, Sharpe/Sortino & Drawdown Recovery): Exact time additivity vs cross-sectional linear weights and IEEE-754 expm1/log1p.Source: QuantCalc Tools Institutional Research DeskLicense: CC-BY-4.0

Reviewing junior quantitative research notebooks reveals one single bug more often than any other: computing daily asset log returns r = np.log(prices / prices.shift(1)), multiplying by portfolio weights r_port = (r @ weights), and summing r_port.cumsum! That one-liner violates conservation of dollars! Why? Because logarithmic returns add cleanly across time (t = 1..T), whereas simple linear returns add cleanly across portfolio constituents (i = 1..N)—and neither return convention is linear across both dimensions simultaneously! Audit exact compounding in our CAGR Calculator, inspect multi-asset risk in our Value-at-Risk Calculator, and consult the CAGR, IRR & Drawdown Hub.

Exact Identities & The Cross-Sectional Jensen Inequality Error#

Let $R{i,t} = (P{i,t} - P{i,t-1}) / P{i,t-1} \in (-1, \infty)$ be asset $i$'s simple discrete return and $r{i,t} = \ln(P{i,t} / P{i,t-1}) = \ln(1 + R{i,t}) \in (-\infty, \infty)$ be its continuously compounded log return ($R{i,t} = e^{r{i,t}} - 1$). Over $T$ time steps for a single asset, telescoping products turns $\ln(PT / P0)$ into an exact sum $\sum{t=1}^T rt$. Conversely, across $N$ assets held at beginning-of-period dollar weights $wi$ ($\sum wi = 1$ with $wi \ge 0$), dollar wealth linearity forces $R{p,t} = \sum{i=1}^N wi R_{i,t}$, meaning exact portfolio log return is the log of a weighted sum of exponentials:

Formula or model from the articleNot independently verified
Exact Time Aggregation vs Cross-Sectional Portfolio Aggregation Laws:
  • 1. Single-Asset Time Aggregation (Exact!) :
  • r{1..T} = ln(PT / P0) = Sum{t=1..T} rt <=> CAGR = exp( (1/Tyrs) * Sum{t=1..T} rt ) - 1
  • 2. Multi-Asset Portfolio Aggregation (Exact!):
  • R{p,t} = Sum{i=1..N} (wi * R{i,t}) <=> r{p,t} = ln( Sum{i=1..N} [ wi * exp(r{i,t}) ] )
  • 3. BUGGY Weighted Log-Return Formula & Exact Error Term (for w_i >= 0):
  • rbuggy = Sum{i=1..N} (wi * r{i,t}) <= r_{p,t} [Strictly understates long-only returns by Jensen's Inequality!]
  • Error = r{p,t} - Sum(wi * r{i,t}) ~= 0.5 * [ Sum(wi * r{i,t}^2) - (Sum wi * r_{i,t})^2 ] (Cross-sectional dispersion!)
  • Worked Proof of the weights @ log_returns Bug (50/50 Portfolio over 1 Period):
  • Asset 1 surges +100% (R1 = +1.00, r1 = ln(2.0) = +0.69315)
  • Asset 2 crashes -50% (R2 = -0.50, r2 = ln(0.5) = -0.69315)
  • True Dollar Wealth ($100 -> $50 in each): $50(2.0) + $50(0.5) = $100 + $25 = $125.00!
  • > Exact Simple Return Rp = +25.00%; Exact Portfolio Log Return rp = ln(1.25) = +0.22314 (+22.31% log)!
  • BUGGY w @ r Code: 0.5(+0.69315) + 0.5(-0.69315) = 0.00000% (Claims portfolio made $0.00—off by -2,500 bps!)

Look at that result: weights @ log_returns told the researcher the portfolio made 0.00% ($100.00), when in reality the account balance sits at $125.00 (+25.00%)! And notice why the second-order error term $\frac{1}{2}[\sum wi ri^2 - (\sum wi ri)^2]$ looks so familiar: missing that cross-sectional dispersion term is literally throwing away the entire Booth-Fama / Shannon's Demon Rebalancing Bonus!

Correct vector pipeline converting log returns to simple returns for cross-sectional portfolio weights and back to log returns for time compounding
Correct vector pipeline converting log returns to simple returns for cross-sectional portfolio weights and back to log returns for time compoundingVerified Exhibit

Even Worse on Short & Leveraged Portfolios: weights @ log_returns Hides Margin Wipeouts!#

For long-only weights ($wi \ge 0$), convexity of $e^x$ means w @ r systematically understates portfolio return. On **short or leveraged portfolios ($wi < 0$ or $\sum w_i > 1$), w @ r commits an infinitely more dangerous error in the opposite direction: suppose you short 100% of a meme stock ($w = -1.0$) that surges $+150\%$ overnight** ($R = +1.50$, $r = \ln(2.5) = +0.9163$).

  • True Account Equity: $1 + (-1.0)(+1.50) = -0.50$ (100% wiped out plus -50% negative-balance margin deficit!).
  • Buggy w @ r Code: $\exp(-1.0 \times 0.9163) - 1 = e^{-0.9163} - 1 =$ -60.0% loss—pretending $40\%$ of your equity survived a trade that bankrupted the account!

Domain Rules: Where to Use Log Returns, Simple Returns & Yield Differences#

Apply Attilio Meucci’s invariance separation across time-series simulation, Markowitz optimization, and fixed-income attribution:

Quantitative Distribution · Use Simple Returns ($Rt = Pt/P_{t-1}-1$)?
Use Simple Returns ($Rt = Pt/P_{t-1}-1$)?
01GARCH / Cholesky Path Simul…Convert via $e^…02Yield / Credit Spread Chang…No ($y_t$ can c…
Quantitative Workflow TaskUse Log Returns ($rt = \ln(Pt/P_{t-1})$)?Use Simple Returns ($Rt = Pt/P_{t-1}-1$)?Mathematical Reason
Multi-Period Time Horizon ScalingYES ($r{T} = \sum{t=1}^T r_t$)No (products skew lognormally)Central Limit Theorem & time additivity
GARCH / Cholesky Path SimulationYES (for log-price assets)Convert via $e^r - 1$ for PnL!Guarantees simulated $PT = P0 e^{\sum r_t} > 0$
Markowitz MVO / Portfolio WeightsNo (breaks dollar additivity!)YES ($\mu_p = w^\top \mathbb{E}[R]$)Single-period wealth is linear in $R_i$
Sharpe, Sortino & Brinson AttributionNo (distorts excess borrow $r_f$)YES ($R{p,t} - r{f,t}$)Margin financing & active weights are linear
Yield / Credit Spread ChangesNO! Use Differences $\Delta y_t$!No ($y_t$ can cross $0\%$ / negative!)Meucci "Quest for Invariance": rates add in bps
Annualized cumulative tracking error in basis points caused by multiplying portfolio weights by log returns across daily, weekly, monthly, and annual frequencies
Annualized cumulative tracking error in basis points caused by multiplying portfolio weights by log returns across daily, weekly, monthly, and annual frequenciesVerified Exhibit

Busting the Myth That "At Daily Frequency, Log and Simple Returns Are Identical!"#

Read the first metric box in the chart above very carefully—this shocks even experienced quants! On one single day, $|rt - Rt| \approx \frac{1}{2}Rt^2 \approx \frac{1}{2}\sigma{\text{daily}}^2$ really is tiny ($\sim 1.9\text{ bps}$ per day for $35\%$-vol stocks). When you sum $252$ daily errors across a year, $252 \times \frac{1}{2}\sigma{\text{daily}}^2 = \frac{1}{2}\sigma{\text{annual}}^2 =$ -612 bps per year! Higher sampling frequency does NOT make cumulative w @ r error vanish, because $N{\text{steps}} \times \mathcal{O}(1/N{\text{steps}}) = \mathcal{O}(1)$!

Applying Attilio Meucci's "Quest for Invariance" Across Asset Classes#

Follow Attilio Meucci's institutional architecture: model equity/FX spot invariants in log space (Cholesky Monte Carlo Engine), model interest rates and credit spreads in basis-point differences ($\Delta y_t$), map every instrument back to exact linear P&L at the horizon before applying weights or computing Cornish-Fisher VaR and arithmetic vs geometric variance drain, and verify production implementations in our methodology and proprietary software stack.

Frequently asked questions

Why doesn’t the difference between daily log returns and daily simple returns vanish over a 1-year backtest?

Because r_t = ln(1 + R_t) = R_t - 0.5*R_t^2 + O(R_t^3) differs from R_t by a SYSTEMATIC one-signed quadratic term (-0.5*R_t^2), not zero-mean noise! Summing -0.5*R_t^2 over 252 trading days accumulates quadratic variation -0.5*sigma_annual^2 regardless of how finely you slice Δt!

How do I convert an annualized log return mean (mu_log) into annualized CAGR?

Simply exponentiate and subtract one: CAGR = exp(mu_log) - 1! In fact, annualized sample mean log return (1 / T_years) * Sum_{t=1..N} r_t equals (1 / T_years) * ln(P_N / P_0), so exp(mean_annual_log_return) - 1 = (P_N / P_0)^(1 / T_years) - 1 === Exact Geometric CAGR!

Why shouldn’t I take log returns of bond yields or interest rates like ln(y_t / y_{t-1})?

Two reasons: first, sovereign yields (EUR, JPY, CHF in 2015–2022) and real TIPS yields frequently drop to 0% or turn negative, crashing ln(y_t); second, bond DV01 P&L scales linearly with basis-point yield CHANGES (y_t - y_{t-1}), not percentage yield ratios!

In numpy/pandas, what is the cleanest vectorized one-liner to compute exact rebalanced portfolio equity from log returns?

Given a (T x N) matrix log_ret and length-N weight vector w: write simple_port = (np.expm1(log_ret) @ w) and then equity_curve = np.exp(np.log1p(simple_port).cumsum) (using expm1 and log1p for full IEEE-754 floating-point precision on small returns!).

Publication record

Published
13 November 2026
Updated
14 November 2026
Author
Elena Rostova, FRM, CQF
Review
Review not recorded
Automated fact check
Passed in the stored pipeline record
Source URLs
4 recorded

AI assistance recorded: anthropic/claude-opus-4.6. This field does not establish human review or source verification.

Academic & Institutional Attribution

Cite This Research & Embed Live Data Card

Open Attribution License (CC BY 4.0)

Standard academic & institutional research paper citation format.

Elena Rostova, FRM, CQF. (2026, November 14). Continuous Log Returns vs Simple Returns: Time Additivity vs Asset Weights. QuantCalc Tools. https://quantcalctools.com/guides/continuous-log-return-time-additivity-vs-simple-portfolio-weights
Explore Cagr Irr Drawdown Compounding Hub (20) →

QuantCalc Tools Data, Benchmarks & Contributor Records

Log Returns vs Simple Returns: Time vs Portfolio Aggregation | QuantCalc Tools