Evaluating Tax-Loss Harvesting with Costs, IRR, and Simulation
Summary
The document considers how to evaluate tax-loss harvesting by comparing an estimated tax benefit with the costs of realizing and reinvesting after a loss. The question proposes expressing the benefit relative to an asset class’s portfolio value and including fee differences, commissions, and bid-ask spread costs, then asks whether internal rate of return and time horizons should also enter the calculation. The response says IRR can be included, but it is not essential to every sound model; in some implementations its inputs may not materially affect the result.
For assessing a strategy, the answer recommends examining historical outcomes and using Monte Carlo simulation to understand the range and likelihood of results. It points to published research that considers tax deferral and differential IRR as an example of a more detailed treatment. The exchange does not provide a complete formula, specify a universal threshold, or quantify the benefits and costs. Its main lesson is that model design varies, and evaluation should account for uncertainty rather than treating a single benefit-minus-cost estimate as conclusive.
Key ideas
- A tax-loss harvesting rule can compare estimated tax benefits with trading and portfolio costs.
- Potential costs include fee differences, commissions, and bid-ask spread losses.
- IRR and time horizons can be modeled, but the answer says they are not mandatory in every implementation.
- Historical analysis and Monte Carlo simulation can help assess strategy outcomes and uncertainty.
- The document does not prescribe a universal threshold or a complete calculation method.
Tags
Full text
# What is an appropriate algorithm to use for tax loss harvesting? # What is an appropriate algorithm to use for tax loss harvesting? I've been reading into how Betterment and Wealthfront have architected their tax loss harvesting algorithms, but they stop short of providing any real examples. Essentially, they both reduce to: ``` Benefit – Cost ≥ Threshold ``` They differ in how they define each term, however. Threshold is proprietary and the result of Monte Carlo research, so let's leave it aside for now. Benefit looks like it is best measured as a percentage of the asset class in the portfolio. For example, if cost basis of the emerging markets fund makes up 10,000 of a portfolio and the loss is $400, does it make sense to say the benefit is ``` ($60 ÷ $10,000) = .6% ``` Then cost, for example, would equal the management fee difference. Let's say .09% and .18%, plus any commissions or bid/ask spread losses expressed as a percent. ``` (.09% - .18%) - 0 - 0 = .09% ``` Making the difference ``` .6% - .09% ≥ Threshold ``` Have I thought about this correctly or am I missing IRR and time horizons? ## Answer by Nathan S. (score 3) https://quant.stackexchange.com/a/16873 The answer is that you may use an approach that includes IRR, but that's not a necessary component of what I would consider a good model. I have seen commercial tools that include them and those that don't. I have also seen practitioners set the variables in packages that include this approach, so that they were not a relevant component of the resulting model. You're going to want both a view of historical results and consider Monte Carlo to get a better feeling for the probability surface of your strategy. This section of the Wealthfront documentation might help you. They show that they have considered approaches incorporating IRR. https://research.wealthfront.com/whitepapers/tax-loss-harvesting/#15a-quantifying_the_value_of_tax_deferral_differential_IRR I don't know who works on Wealthfront's algorithms, but I would like to. I cannot speak to the quality of their services in any way, but I can say that the research in the whitepaper above is not naive. And that is saying something special in the world of tax efficient portfolio management for retail investors.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.