Presets
Playback
Market (two assets)
Illustrative returns and risks of a stock fund and a bond fund.
Rebalancing policies
Target weight, band width, calendar period and the cost of one trade.
Time
Shocks and fast-forward
Test the policies against a sudden move.
Camera
Quality
Layers
Model checks
The science
Two assets, one random path
Stocks and bonds follow correlated geometric Brownian motion, dS/S = mu dt + sigma dW, simulated daily (252 trading days a year). Holdings are tracked in value, so without trades the stock weight is w = h1 / (h1 + h2). The returns, volatilities and correlation are illustrative round numbers, not forecasts.
Why weights drift
Let Z = ln(S1/S2). It is Brownian motion with drift nu = (mu1 - s1^2/2) - (mu2 - s2^2/2) and variance s^2 = s1^2 + s2^2 - 2 rho s1 s2 per year. A buy-and-hold weight is exactly w = logistic(logit(w0) + Z), so the log-odds of the weight is a random walk. A bigger stock volatility, a lower correlation or a larger return gap makes the weight leave the target faster.
Three policies
Buy and hold never trades. Calendar sells the overweight asset and buys the other back to the target on a fixed schedule (every P months, 21 trading days per month). Threshold band trades only when |w - w0| exceeds the half-width, then returns to the target. A trade moves 2 |w - w0| W of value and costs c basis points of it (1 bp = 0.01 %), so fee = c x 1e-4 x 2 |w - w0| W, taken out of wealth before the new holdings are set.
Tracking error and cost
Tracking error here is the root-mean-square gap between the daily weight and the target, in percentage points. The cost-drag axis is the total fees divided by wealth, per year, in basis points. The frontier graph reruns 150 random 10-year paths for 7 calendar periods and 7 bands on the same paths and plots both curves; the white rings are your current settings.
When does a band trade? (closed form)
Starting at the target, the log-odds Z must reach u = logit(w0 + band) - logit(w0) or l = logit(w0 - band) - logit(w0). For Brownian motion with drift nu the chance of leaving through the top is P_u = (e^(-theta l) - 1) / (e^(-theta l) - e^(-theta u)) with theta = 2 nu / s^2, and by Wald's identity the mean exit time is E[tau] = (u P_u + l (1 - P_u)) / nu, or -l u / s^2 when nu = 0. With daily monitoring both barriers are shifted outward by 0.5826 s sqrt(dt) (Broadie-Glasserman-Kou correction); the page shows that this matters.
Rebalancing premium
A constant-mix portfolio has log growth g = w mu1 + (1 - w) mu2 - sigma_p^2 / 2 with sigma_p^2 = w^2 s1^2 + (1 - w)^2 s2^2 + 2 w (1 - w) rho s1 s2. This is higher than the value-weighted average of the assets' own log growth rates by (w s1^2 + (1 - w) s2^2 - sigma_p^2) / 2, the so-called rebalancing (volatility-harvesting) premium. It is small, and fees can easily eat it.
Simplifications
Constant volatilities and correlation, no jumps, no autocorrelation or momentum, no taxes, no bid-ask spread beyond a proportional fee, no minimum lot sizes, continuous cash availability, and trades at the closing price. The beam tilt and the coin counts show the weight gap relative to target, not a physical mass balance. This is an educational model and not investment advice.
What to try
- Press Run 10 years now on the default preset and compare the three weight lines: the grey buy-and-hold weight wanders far from the green target while the others stay close to it.
- Choose Tight band 1 pp: tracking error is tiny but the band policy pays far more in fees than the calendar. Then choose Expensive trading and watch the fee jars fill.
- Choose Free trading: with zero cost, trading often costs nothing and the tightest control wins.
- Press Stocks crash 30 % and see which policy reacts first: the band trades the same day, the calendar waits for its next date.
- Read the frontier graph: at the same tracking error, which curve (calendar or band) has the lower fee drag in your market?
- Compare the histogram in the last graph with the formula line: the mean time to breach the band is not a guess, it follows from the log-odds drift.