Trying to model interacting retirement decisions (SS timing, withdrawals, Roth conversions) — does this approach make sense?
I’m starting retirement this year and have been going pretty deep into the planning side.
One thing that stood out is how interdependent a lot of the decisions are:
- Social Security timing
- withdrawal order across taxable / tax-deferred / Roth
- whether and how aggressively to do Roth conversions before RMDs
Most of the tools I tried, including ones like Boldin and ProjectionLab, felt very scenario-based — you tweak one variable at a time and compare outcomes. The problem I kept running into is that the “best” choice for one decision often depends on what you assume for the others.
So I started building something for myself to explore this more systematically.
The way I’ve approached it so far:
- use Monte Carlo to answer “does this plan hold up?” type questions
- then run a broader search across combinations of decisions (SS timing, withdrawal order, Roth conversion strategies) to try to find more efficient outcomes
For a typical couple, this ends up evaluating a pretty large number of combinations rather than relying on a few hand-picked scenarios.
Before I go too far down this path, I’d really appreciate input from people here:
- Do you think this kind of “search across strategies” is actually useful, or does it risk overfitting to assumptions?
- How are you currently handling these interacting decisions?
- Are there tools or workflows you feel already solve this well?
Would especially value critical feedback.
EDIT: Really appreciate all the thoughtful responses — especially the references to tools like Pralana and Boldin, and the discussion around rules of thumb vs. optimization.
One thing I’m still trying to understand better is whether systematically exploring combinations of decisions actually leads to meaningfully different outcomes, or just adds complexity on top of what people are already doing.
I wrote up how I’ve been thinking about this in more detail here: https://pinebrook.ai/prisma/approach.html
Would especially appreciate any pushback on whether this feels genuinely different, or just a more exhaustive version of existing approaches.