RSJ Portfolio Portfolio Optimization

← All posts

2026-11-10

How to Verify a Portfolio Optimizer's Allocation Results

How to Verify a Portfolio Optimizer's Allocation Results
portfolio optimization modern portfolio theory portfolio correlation portfolio optimization hierachical risk parity minimize volatility minimize semivariance maximize sharpe

How to Verify a Portfolio Optimizer’s Allocation Results

Every portfolio optimizer returns weights that sum to 100%. That is exactly what makes it dangerous: the arithmetic is always correct, even when the inputs, the settings, or the interpretation are wrong. Knowing how to verify a portfolio optimizer’s results is the difference between a plausible-looking table and an allocation you can actually put capital behind.

This article is a verification checklist. It walks through what a trustworthy run should leave behind, how to audit each artifact in RSJ Portfolio, and how to read the results that look strange at first glance. It applies to any optimizer, but the concrete steps use RSJ Portfolio’s screens and user guide, because a checklist you can point at is worth more than a principle.

Why Verification Matters Before You Trust the Weights

An optimizer is a function: inputs in, weights out. If a single input is stale or a single constraint is looser than you intended, the output still looks like a portfolio. Nothing in the weight table flags the mistake.

The cost of skipping verification is usually one of four things:

  • Silent data errors. A price series that ends a month early, a symbol that changed ticker, a corporate action that the cleaning pipeline handled differently than you assumed. The correlation matrix is built from whatever the pipeline produced.
  • Mis-set constraints. A per-asset cap you meant to tighten, or a floor you inherited from a previous run. Constraints shape the feasible set, so a wrong constraint produces a wrong — but perfectly formatted — allocation.
  • Objective mismatch. Minimum semivariance and maximum Sharpe ratio answer different questions. If the objective does not match how the investor thinks about risk, the weights are correct and useless at the same time. We cover the trade-off in Maximum Sharpe Ratio vs Minimum Semivariance Allocation.
  • Currency drift. A portfolio quoted across several currencies inherits the conversion path. If the local currency setting is not the one you meant, the risk estimates shift with it.

Treat verification as a repeatable checklist, not a one-off sanity check performed the first time and never again. Run it every time the candidate list, the date range, or the objective changes. The configuration questions that matter most before your first run are covered separately in First Run Setup: Portfolio Optimizer Configuration That Matters; this article assumes those choices are already made and asks whether the run you just produced actually used them.

What a Verifiable Result Looks Like

A verifiable result is one you can reconstruct without asking anyone what happened. Six artifacts make that possible:

  1. The candidate list — the exact symbols that entered the optimization, frozen at a point in time.
  2. The objective — minimum volatility, maximum Sharpe ratio, minimum semi variance, hierarchical risk parity, or conditional value at risk.
  3. The risk model and return model — which estimator turned the price history into numbers the optimizer could minimize or maximize.
  4. The weight constraints — the bounds each position had to respect.
  5. The weight vector — the allocation itself.
  6. The data window — the data provider, the cleaning settings, the local currency, and the date range.

Items 1 through 5 explain what the optimizer decided. Item 6 explains what it was looking at. Without item 6, a weight vector is a number with no context; you cannot re-run it, and you cannot audit it.

That is the real dividing line. A black-box result is a set of weights someone handed you. A verifiable result is a set of weights plus the configuration that produced them, such that a colleague could reproduce the run on their own machine and get the same answer.

In RSJ Portfolio, the Recommendation view is where the optimizer’s parameters and the resulting allocation appear together, which makes it the natural starting point for any audit. The Configuration screen holds the data-side settings, and Selection holds the candidate list. Three screens, one run.

Step-by-Step: Verifying a Run in RSJ Portfolio

Step 1 — Confirm the data pipeline

Open the Configuration screen and check four things before you look at a single weight:

  • The data provider you configured.
  • The price cleaning settings applied to the raw series.
  • The local currency for the portfolio.
  • The date range used for the estimate.

The Data and currencies page documents the price cleaning pipeline, the local cache, and the conversion path for portfolios quoted in several currencies — read it once, then use it as the reference when a number surprises you. The point of this step is to confirm that what the optimizer consumed matches what you believe you asked for.

Step 2 — Inspect the candidate list

Switch to Selection and read the symbols as a list, not as a count. You are looking for two failure modes:

  • Unintended members. A symbol added during exploration and never removed; a proxy for an asset you cannot hold directly.
  • Stale members. A ticker with a short history, a series that stops before the end of your date range, or a symbol whose name no longer matches the instrument.

A candidate list of 40 names where three have six months of history is not a 40-name universe. It is a 37-name universe plus noise.

Step 3 — Read the optimizer settings

In the Recommendation view, confirm the three settings that define the problem:

  • Objective. Minimum volatility, maximum Sharpe ratio, minimum semi variance, hierarchical risk parity, or conditional value at risk — whichever you intended, not whichever was left selected.
  • Risk model. RSJ Portfolio ships eight risk models plus selectable return models. The choice changes the estimate of risk, and the allocation follows the estimate.
  • Weight constraints. Per-asset bounds and any other limits you set.

If you cannot state the objective, the risk model, and the constraints in one sentence each, you are not ready to read the weights.

Step 4 — Cross-check the weights

Now the arithmetic. Three checks, in this order:

import json
from pathlib import Path

# weights as exported from the Recommendation view
weights = json.loads(Path("allocation-2026-11-10.json").read_text())

total = sum(weights.values())
assert abs(total - 1.0) < 1e-6, f"weights sum to {total:.6f}"

cap, floor = 0.10, 0.00
violations = {s: w for s, w in weights.items() if not (floor <= w <= cap + 1e-9)}
print("constraint violations:", violations or "none")

top = sorted(weights.items(), key=lambda kv: kv[1], reverse=True)[:5]
for symbol, weight in top:
    print(f"{symbol:<8} {weight:>7.2%}")

First, the sum. Second, the constraints. Third, the shape of the distribution — the five largest positions tell you more about the run than the smallest thirty-five.

Then ask the qualitative question: is the correlation rationale visible? The method allocates toward candidates that correlate weakly with each other, or whose downside correlates weakly. If two of your largest positions are near-duplicates of the same exposure, either the correlation estimate disagrees with your intuition or something upstream is off. That disagreement is worth investigating either way.

Step 5 — Re-run with a perturbation

Change one thing and re-run: shift the date range by a few months, or switch the risk model. Compare the two weight vectors.

# keep each run's export in its own directory
ls verification/
# 2026-11-10-a/  allocation.json  config.md
# 2026-11-10-b/  allocation.json  config.md

A small shift is reassuring. A large swing is a signal to investigate — not automatically a bug. It may mean the allocation is sensitive to a handful of observations, that the date range is too short for the number of candidates, or that the risk model choice is doing more work than you thought. Overfitting is the deeper version of this question, and it has its own checklist in How to Check That a Portfolio Optimizer Is Not Overfitting.

Interpreting Results and Handling Edge Cases

Three outcomes look wrong to first-time users and are usually informative rather than broken.

A single asset dominates. Read it as a statement about the correlation structure and the constraints. A dominant weight can be a legitimate consequence of one asset being weakly correlated with everything else, or a sign that a cap is too loose or the date range too short. Re-run with a tighter constraint and with a different window. If the dominance survives both, it is a property of the data.

Weights pinned at a boundary. Positions sitting exactly at a cap or floor mean the unconstrained optimum wanted to go further. That is information about the constraint, not a failure of the optimizer. Loosen the bound in a scratch run to see how far it would have gone, then decide which version you can live with.

A nearly equal-weight solution. Often the honest answer when the candidate set is homogeneous or the history is short: the optimizer cannot distinguish the assets, so it spreads. A near-equal-weight result on a list of twenty large-cap peers is a finding, not an error.

The risk estimator deserves its own line here. Switching between the eight risk models changes the allocation because each model estimates risk differently — that is the intended behaviour, not instability. Verify that the selected model matches how the investor actually thinks about risk, and use the Concept chapter of the user guide for the statistical background. For downside-specific interpretation, see Semivariance Explained: Measuring Only the Losses That Hurt.

Documenting the Verification for Future Reference

The verification is only worth the run if you can repeat it. Save the configuration, the candidate list, and the resulting weights together, in one directory per run, and write down the perturbation you tried.

A lightweight log is enough:

run_id: 2026-11-10-a
objective: minimum semi variance
risk_model: "<as selected in the Recommendation view>"
return_model: "<as selected>"
constraints:
  max_weight: 0.10
  min_weight: 0.00
data:
  provider: "<configured provider>"
  local_currency: EUR
  date_range: 2019-01-01..2026-10-31
  cleaning: "<settings as shown on the Configuration screen>"
checks:
  weights_sum_to_one: true
  largest_position: 0.087
  perturbation: "window shifted 6 months -> max |delta w| = 0.021"
anomalies: []

RSJ Portfolio’s exports are the practical way to capture the allocation for review or to send the run to a colleague — export the weights, keep the config alongside them, and the run becomes auditable months later.

A documented trail is also your first defence against hindsight. A result that only looks good because you remember the version of the run that worked is not a result; the log is what stops that story from being told. The overfitting checklist linked above builds directly on this habit.

FAQ

How do I know the optimizer used the data I intended? Check the Configuration screen in RSJ Portfolio: it shows the data provider, the price cleaning settings, the local currency, and the date range. The Data and currencies page in the user guide documents how the cleaning pipeline and multi-currency conversion work, so you can confirm the inputs match your intent before reading the weights.

Why do the weights change when I switch the risk model? Each of the eight risk models estimates risk differently, so the optimizer’s allocation responds to that estimate. A change in weights is expected; verify that the model you selected reflects how you think about risk, and use the Concept chapter of the user guide to understand the statistical background.

What should I do if a single stock dominates the allocation? First confirm the candidate list and constraints are what you intended. A dominant weight can be a legitimate result of the correlation structure, or a sign that a constraint is too loose or the data range is too short. Re-run with a different date range or a tighter constraint to see whether the dominance persists.

Can I reproduce a run later to verify it? Yes, if you save the configuration (objective, risk model, constraints, data provider, date range) alongside the resulting weights. RSJ Portfolio’s exports let you capture the allocation, and the user guide’s Recommendation chapter explains the parameters and results in detail.

How long should verification take? For a run whose configuration you already know, the five steps above take a few minutes — most of the time goes into step 2, reading the candidate list properly. The first run after a configuration change takes longer, because steps 1 and 3 require re-reading rather than confirming.

Does verification differ by objective? The artifacts are identical, so the checklist does not change. What changes is step 5: a minimum-volatility run tends to move less under a shifted window than a maximum-Sharpe run, which chases return estimates as well as covariance. Expect more sensitivity, and do not read it as instability until you have compared two windows.

What if a re-run produces different weights? That is the question verification exists to answer. If the only difference is the perturbation you introduced in step 5, the difference is the finding. If the configuration is identical and the weights still differ, check that the candidate list, date range, and data provider are truly the same before concluding anything else.

Conclusion

Verification is not a vote of no confidence in the optimizer — it is the step that turns an output into an argument. Confirm the data pipeline, read the candidate list, state the objective and risk model out loud, cross-check the weights against the constraints, then perturb the run once to see how much the answer moves. Save what you did.

RSJ Portfolio is built to make that audit short: the configuration, the selection, and the resulting allocation live on separate screens that map directly onto the checklist, with exports to capture each run. If you have not started yet, begin with the overview and the getting started guide; if you have a run open right now, go to the Recommendation view and start at Step 3.

Related posts

EU label: AI-generated content