USE CASE
Minimizing portfolio volatility for AI agents
A person tells their agent the portfolio feels too risky. Abacus Labs turns that into a target the risk model can measure, returns options that cut volatility by different amounts, and prices each one in tax, turnover and what the person gives up.
The problem
"Less risky" has no number in it. Volatility is a property of the whole portfolio, driven by how positions move together, not a list of tickers with a safe reputation. Two defensive stocks can share the same factor and fall on the same day. Estimating a covariance matrix for a few hundred names from price history alone needs more history than exists, which is why professional tools use factor models.
The payoff from doing it properly is well documented. Clarke, de Silva and Thorley found that a long-only minimum-variance portfolio of large US stocks carried about three quarters of the realized risk of the capitalization-weighted market over 1968 to 2005 (Journal of Portfolio Management, 2006). Reaching it in a real account means selling positions with gains, so every option has a tax price.
A language model alone names "low volatility" tickers from memory. It has no covariance, no lots and no way to tell the person how much risk actually came off.
What the agent asks
USER This swings too much. Can you calm it down?
agent -> analyze(account)
agent -> propose(account, "cut volatility by about a quarter, keep taxes low", options=3)
AGENT Your account moves about 31% a year; the broad market
moves about 17%. Most of that is NVDA plus three
other chip names moving together. Three ways down:
1 trim the chip names vol 31% -> 24% tax $2,100
2 add bonds and broad ETFs, sell nothing
vol 31% -> 22% tax $0
3 move half to a broad ETF core vol 31% -> 20% tax $4,800
USER What if I only add, nothing sold?
agent -> propose(option_2, change={ "amount": 15000 })
USER Do that.
agent -> commit(option_2b)
What Abacus Labs returns
{
"status": "ok",
"trades": [
{ "symbol": "BND", "side": "buy", "qty": 120 },
{ "symbol": "VTI", "side": "buy", "qty": 18 },
{ "symbol": "XLV", "side": "buy", "qty": 25 }
],
"risk": {
"volatility": { "before": 0.31, "after": 0.22, "benchmark": 0.17 },
"beta": { "before": 1.42, "after": 1.08 },
"top_factor_exposures_after": { "technology": 0.38, "momentum": 0.21, "size": -0.12 }
},
"tax": { "realized_gain": 0.00, "realized_loss": 0.00 },
"checks": { "cash": "pass", "whole_shares": "pass",
"wash_sale": "pass", "limits": "pass" },
"explain": "Adds bonds, a broad index fund and healthcare with the new cash. Sells nothing, so no tax. Cuts volatility by about 30% because the new positions move little with the chip names.",
"certificate": "cert_2026-11-06_c3e9"
}
How it works
Our point-in-time risk model gives every US stock and ETF a set of factor exposures and a specific risk. From these Abacus Labs computes portfolio volatility, beta and the share of risk each position contributes. The goal "about a quarter less" becomes a volatility target. Constraints keep the plan realistic: a tax budget, no new names if asked, sector caps, whole shares and a cash floor.
Options differ in the lever they pull: sell the names that contribute most risk, add positions that move little with the rest, or shift part of the account to a broad core. Explore ranks them in tens of milliseconds. Commit re-solves the chosen option with an exact mixed-integer solver and returns the certificate. The learned explore step is built to answer about 10x faster than a conventional solver run, which is what lets the agent answer "what if I only add" inside the conversation.
The risk model covers US equities and ETFs and is evaluated out of sample from 1993 to 2026. Abacus Labs does not forecast which option will earn more. It reports how much risk each one carries and what it costs to get there.
Annualized volatility, example account
Current ######################################## 31% 1 trim chips ############################### 24% 2 add, no sales ############################ 22% 3 broad ETF core ########################## 20% Broad market ###################### 17%
Example account, illustrative numbers from the conversation above. Volatility is the risk model's estimate, not a realized return.
Who uses this
- General assistants via MCP
- The first thing people ask about a linked account is whether it is too risky. The MCP server answers with numbers the assistant can relay.
- Broker agents
- Risk-level requests such as "moderate" or "steady" become measured targets instead of a model portfolio.
- Advisor tools
- Matching an account to a client's stated risk tolerance, with the tax cost of getting there shown up front.
The Personal plan is free: 3 portfolios and 300 credits a month, no card. Paid plans add portfolios, credits and client use.