Six months ago we argued that investors don't need more information, they need conviction. Since then, the engine behind Finxer.ai has grown from a handful of opinions into a full pipeline that turns those opinions into a portfolio. This post walks through how it works, step by step.
1. Many agents, many philosophies
Every asset in our universe is evaluated independently by several agents. Each one replicates the philosophy of a renowned investor: a Warren Buffett agent looks for durable moats, consistent earnings and rational management; a Benjamin Graham agent insists on a strong balance sheet and a deep margin of safety. They read the same financial statements, but they ask very different questions.
Each agent produces a clear verdict (bullish, neutral or bearish) together with a confidence level and the reasoning behind it. A synthesis agent then weighs these perspectives into a single house view per asset. You can read every one of these analyses, every week, in our research section.
2. Starting from what the market already believes
Opinions alone don't make a portfolio. Naively buying everything the agents like leads to concentrated, unstable allocations that swing wildly with every new signal. So we don't start from the agents. We start from the market.
Using a Bayesian approach in the spirit of the Black–Litterman framework, we first derive the equilibrium expected returns: the returns implied by current market capitalizations, assuming the market as a whole is in balance. This becomes our prior belief. Without any opinion at all, the model would simply hold something close to the market.
The market is the prior. The agents are the evidence.
3. Updating beliefs with confidence-weighted views
The agents' evaluations then enter as views on each asset. Crucially, each view carries its own uncertainty: a high-conviction bullish call from several agreeing agents moves the model a lot, while a hesitant, split verdict barely moves it at all.
Combining the prior with these views yields posterior expected returns: a blend of what the market implies and what our agents believe, weighted by how confident they are. The result is a portfolio that is opinionated where there is conviction and anchored to the market where there isn't, and that is typically more diversified and more stable over time than one built on signals alone.
- Disciplined by default. Weak or conflicting opinions leave allocations close to equilibrium.
- Bold when it matters. Strong, consistent conviction is allowed to tilt the portfolio meaningfully.
- Tested every week. Key parameters, such as risk aversion and the strength given to the views, are backtested across a grid, and the most robust risk-adjusted configuration is selected. You can explore that grid on the portfolio page.
4. Dynamic allocation: protecting principal first
A good equity portfolio is only half of the answer. The final layer, and the focus of our next iteration, is a dynamic allocation mechanism that decides how much capital should be exposed to that portfolio at all, and how much should sit in safe, fixed-income assets.
When conditions are favorable, the mechanism leans into the risky portfolio to capture opportunities. When volatility rises or conviction fades, it shifts toward safety. The design goal is simple: take risk when it is rewarded, and prioritize the preservation of principal when it isn't.
What comes next
Risk-aware portfolios are live today, and the Dynamic Allocation Engine is our next major milestone, scheduled for the end of 2026. As always, you can follow our progress on the roadmap.
As with everything on Finxer.ai, nothing here is investment advice. Models can be wrong, backtests are not guarantees, and investing always involves the risk of losing capital.