In thirty seconds Arithmos goes from a sentence to a 40-stock basket with a documented selection rule, transparent weights, and a 25-year backtest. The value isn’t the speed. It’s that every decision the agent made is written somewhere you can read — and disagree with — before you execute.
Why plain English in the first place
The standard workflow for a custom index is: figure out a thesis, pick a universe, write a screen in a notebook, handle survivorship bias, handle rebalance math, write a backtest, review, iterate. It’s a week of work for someone who already knows the tools. For most people it’s just inaccessible — which is why “direct indexing” has been stuck as a family-office product for forty years.
A prompt-to-portfolio system compresses that week into a sentence. The rule is still in charge — the agent just writes it for you and shows you the code. Think of it as a junior quant who reads well, writes cleanly, and never gets tired.
What the agent actually does
When you paste a prompt, the agent works through a fixed pipeline. Every step is a real tool call — not a guess:
- Parse the thesis. What is the user asking for — a sector, a factor, a thematic, a values screen? The agent produces a structured intent it can reason about.
- Pick a universe. US large-cap, global developed, a specific NAICS branch — whichever is widest enough to cover the thesis without importing noise.
- Screen the universe. Apply fundamental, factor, and thematic filters. The agent uses real market- data APIs — same data an institutional quant would screen against — not its pre-training.
- Score and rank. For thesis-driven prompts, each surviving name gets a score that later drives weights.
- Weight and cap. Equal-weight by default, with an 8% per-name cap; factor-weighted when the prompt calls for it.
- Finalize.Emit a structured output (the holdings, the weights, the rationale) that’s validated against a strict schema before any of it hits the UI.
- Backtest. 25 years against SPY with honest survivorship and rebalance handling, deterministic and cached for anyone who views the index later.
Example: one sentence, one index
What the agent emits:
- Universe:US-listed mid/large-cap software (NAICS 5112 + 5415) with market cap > $500m.
- Selection rule:positive free cash flow over the trailing twelve months, gross margin > 60%, EV/ Sales < 12×, Rule-of-40 score > 30.
- Result: ~35 names.
- Weights:equal-weight, 8% per-name cap (doesn’t bind at this size).
- Rebalance: quarterly.
- Backtest: annualised return, Sharpe, max drawdown, rolling alpha vs SPY — all visible with the index.
You can now do the thing that matters most: read the rule, decide whether you agree with it, edit, or regenerate.
Guardrails — what stops the agent going off
The interesting part of building a prompt-to-portfolio system isn’t the LLM. It’s the guardrails that keep it honest:
- Strict output schema.The final holdings tool’s arguments are validated against a Zod schema. Invalid weights, missing tickers, rogue sectors — the agent is forced to retry.
- Real data, not model memory.The agent never “knows” a market cap — it calls the market-data provider. Training-data staleness is structurally impossible for anything time-sensitive.
- Bounded tool budget.Capped iterations and capped tool calls so a prompt can’t spiral.
- Deterministic backtest. Same input, same output — the backtest engine is pure and runs outside the LLM loop entirely.
- Execution is a separate step. The agent never places orders. Execution is a human-approved, dry- runnable action in the product.
Compared to a chatbot that picks stocks
You can ask any chatbot “which stocks should I buy?” — that’s not the same product.
- A chatbot picks stocks from its pre-training (months out of date), can’t show you its selection rule, and can’t reproduce its answer two days later.
- Arithmos picks a rule, runs it against today’s data, shows you the rule in plain English and the rationale for each decision. Running it tomorrow on the same prompt produces a closely-related answer because the ruleis stable — the holdings just reflect today’s universe.
That’s the difference between “I think these are good” and “here’s the rule, here’s the backtest, here’s how often you rebalance.”
Reading the rule is the whole point
The single most valuable artefact the agent produces is the one-paragraph selection rule at the top of each index. Read it. If you can’t defend it on a page, don’t invest behind it — regenerate the prompt or edit the inputs.
Prompts that work (and don’t)
Prompts that produce strong indices
- “Profitable US software businesses at reasonable valuations.”
- “European defence primes and their top-tier suppliers, equal-weighted, quarterly rebalance.”
- “Dividend aristocrats with debt-to-equity under 1 and payout ratio under 70%.”
- “Global AI infrastructure — semiconductors, hyperscale cloud, and networking — ex-China.”
Prompts that need sharpening
- “Good companies.” — define good. The agent has to guess, which means you will disagree with it.
- “Stocks that will beat the market.” — this is a wish, not a rule.
- “Warren Buffett-style.” — specific enough to hand the agent (it will screen for quality + value + low leverage), but the agent will do better if you name the tests directly.
FAQ
Is the agent making up the financials?
No. Every number the agent uses comes from live market-data providers at the moment of generation. The LLM’s job is to choose which numbers to look up, not to invent them.
Can two users get the same index from the same prompt?
Almost identical. The selection rule is stable; the universe is today’s data. Re-running a week later will produce a very similar basket with small drift — exactly what a quarterly-rebalanced rule should do.
How do I edit a generated index?
Every generated index has an edit panel — you can exclude individual names, tweak weights, change the cap, or rewrite the prompt and regenerate. The rule stays human-readable the whole way through.
Is this legal advice / regulated financial advice?
No. Arithmos is a tool for constructing rules-based portfolios. The rule is yours; the decision to invest behind it is yours. Nothing on the site is personalised investment advice.