Portfolio research, ML signals, and human-approved execution — automated end to end.
What it is
What is the AI Investment Research Engine?
The Investment Research Engine is a multi-agent system that connects directly to your portfolio via Plaid and Yodlee, runs continuous news and sentiment research, generates entry and exit signals using XGBoost and LSTM models, and presents human-approved orders to your broker — all without manual data pulling or spreadsheet work.
Think of it as a research analyst and risk manager working in parallel, 24 hours a day. It reads the news, models the signal, calculates the position size, and brings you a draft order — while you stay in control of every execution decision. Nothing trades without your approval.
This matters for compliance: the engine generates signals and draft orders, but every trade requires human sign-off before routing to a broker. Every decision is logged with the timestamp and approver identity — which is what institutional risk and compliance frameworks require, whether you are an SEC-registered RIA or a family office.
How it works
How the Investment Research Engine works, step by step
Six agents work in sequence — data always flows from live sources, never from a cached spreadsheet. Each agent is specialised: the LSTM that models price signals is different from the Kelly calculator that sizes positions. Here is exactly what happens at each step.
- 01
Data Ingest
Connects via Plaid and Yodlee to every linked brokerage and bank account. Pulls current holdings, transaction history, cost basis, and real-time balances — no manual export, no CSV upload. The engine always starts with live portfolio state, so every signal and risk metric is calculated against what you actually hold right now.
What you get A complete, live picture of the portfolio before any research or analysis begins.
- Plaid
- Yodlee
- IBKR API
- Schwab API
- Fidelity API
- 02
Market Research
A research agent scans news, earnings call transcripts, SEC filings, and analyst commentary across every held position and watchlist ticker. Sentiment is quantified on a rolling basis and signals are calibrated to your specific holdings — not generic market headlines, but events that historically precede price moves in the securities you own.
What you get A prioritised signal feed — what matters to your portfolio right now, ranked by conviction.
- News APIs
- SEC EDGAR
- Earnings transcripts
- Analyst feeds
- Sentiment model
- 03
Signal Generation
Gradient boosting (XGBoost) and LSTM neural networks trained on 10 years of price, volume, and sentiment data produce entry and exit signals with a confidence score per signal. The Kelly criterion sizes each position so you are never overexposed. Signals are probabilistic decision inputs, not instructions — every one is human-reviewed before any order is placed.
What you get A ranked list of actionable signals — entry price, exit target, position size, confidence score.
- XGBoost
- LSTM
- Regression ensemble
- Kelly criterion
- 10-year backtest
- 04
Order Execution
Every signal generates a draft order — ticker, direction, size, order type. Nothing routes to the broker until a human approves it. Once approved, the order executes via your broker API. Partial fills, rejections, and fill confirmations are logged and fed back into the portfolio state for the monitoring step.
What you get Approved orders placed in seconds, with a full audit trail of who approved what and when.
- IBKR API
- Alpaca
- Schwab API
- TD Ameritrade
- Human approval gate
- 05
Portfolio Monitor
Continuous monitoring of portfolio-level risk metrics — beta, Sharpe ratio, max drawdown, sector concentration, and correlation. Threshold breaches trigger immediate alerts. The monitor feeds current portfolio state back into the signal model, so every research cycle starts with live data — not what the portfolio looked like when the market opened.
What you get Live risk visibility — you know the portfolio's health at every moment, not just when you log in.
- Real-time P&L
- Beta tracker
- Drawdown alert
- Sharpe calculator
- Sector exposure
- 06
Rebalance & Report
When sector or position drift exceeds the configured threshold, the rebalance engine calculates the minimum set of trades to restore target allocation. Tax-loss harvesting opportunities are flagged automatically. The rebalance plan is presented for approval — not executed automatically — and the resulting report is formatted for client delivery.
What you get A rebalance plan that is tax-aware, minimal in turnover, and client-ready — without a spreadsheet.
- Drift detection
- Tax-loss harvesting
- Rebalance scheduler
- Client report
The problem
The research problem it solves
Portfolio managers and research analysts at mid-size investment firms face a structural problem: the tools exist to do quantitative research, but the data pipeline between market sources and model inputs is entirely manual.
- Pulling portfolio data from multiple brokerages into a single view takes 1–2 hours per morning before any analysis begins.
- News and sentiment monitoring is ad hoc — the analyst reads what they happen to see, not a systematic signal scan calibrated to their holdings.
- ML models exist but sit idle because re-running them requires a manual data refresh and export cycle.
- Rebalance calculations live in Excel — not tax-aware, not version-controlled, and one formula error away from a costly mistake.
- Trade execution is disconnected from the model output: the signal lives in one tool, the order is placed in another by hand.
- Portfolio-level risk metrics — beta, Sharpe, drawdown — are checked periodically, not continuously.
The engine does not replace the portfolio manager — it removes the hours between signal and action, so the manager spends time on judgment, not data plumbing.
Time to value
How fast you go live
Most teams are live in 2–3 weeks.
- Week 1Connect data sourcesAuthenticate Plaid and Yodlee to your brokerage accounts. Map holdings and cost basis. Validate data against your own records before any model runs.
- Week 1–2Configure signal modelLoad your watchlist and portfolio. Set risk parameters — sector limits, max position size, drawdown thresholds. Run the first backtest against your actual holdings on 10 years of historical data.
- Week 2Approval gate + executionConnect to your broker API. Wire the human approval gate. Run 10 live signals through the approval flow before any live orders are placed.
- Week 2–3Monitor and calibrateRun live for two weeks with daily review. Adjust signal thresholds and Kelly fraction based on observed performance before full deployment.
What you need to start
- Portfolio held at any major US brokerage (IBKR, Schwab, Fidelity, TD Ameritrade, Alpaca) — Plaid-connected.
- A defined investment policy — sector limits, maximum position size, drawdown tolerance.
- A designated approver for orders — PM, CIO, or compliance officer.
- Historical portfolio data for the backtest — prior-year statements or a brokerage export.
No Bloomberg terminal required. The engine works with publicly available market data feeds and your brokerage's own API — tools you already have access to.
ROI
The return on an Investment Research Engine
For a mid-size RIA or family office managing $50M–$500M, the compounding value is time: two hours per day returned to the portfolio manager means 500 hours per year redirected from data pulling to judgment calls and client relationships. On a $100M portfolio, one well-timed signal identified by the model and acted on in seconds — instead of spotted an hour later — can represent more than the engine's annual cost. The model is backtested against your actual holdings before go-live, so you see the historical Sharpe and win rate before committing to live capital.
Proof
What investment teams say
“We were running three separate spreadsheets just to get a clean morning view of our positions. The engine replaced all of that — live data, signals ready by 8 AM, and we haven't touched Excel for portfolio data in four months.”
“The XGBoost model surfaced a re-entry signal three hours before I would have manually identified it. That one trade justified the first quarter's cost.”
“The approval gate was non-negotiable for us — our compliance requires every trade decision to be human-authorised and logged. The audit trail the engine produces has simplified our quarterly compliance review.”
FAQ
Investment Research Engine FAQ
Does the engine place trades automatically?
No. Every signal generates a draft order, and nothing routes to the broker until a human approves it. The approval can be one click in a dashboard or a Slack message — your choice. The audit log records who approved, at what time, and what the signal was. This is a hard design constraint, not an option — it is what institutional compliance requires.
Which brokerages does it connect to?
Any brokerage supported by Plaid or Yodlee for portfolio data — which covers the major US custodians including IBKR, Schwab, Fidelity, TD Ameritrade, and Alpaca. Order execution connects via the broker's own API.
How is the ML model trained and validated?
XGBoost and LSTM models are trained on 10 years of historical price, volume, and sentiment data. Before go-live, we run a full backtest against your actual holdings and watchlist — you see the historical Sharpe ratio, win rate, and max drawdown the signal model would have produced. You validate the model before it runs on live capital.
Is this suitable for a registered investment advisor?
Yes. The engine is designed for SEC-registered advisors. Every trade decision is human-authorised and audit-logged. The engine does not provide investment advice; it generates signals as decision inputs for the PM. The PM retains full discretion and accountability — which is what Form ADV and GIPS compliance require.
What happens if the market moves against a signal?
Stop-loss levels are built into every signal output. If a position moves against the entry by the configured stop, the monitor triggers an alert and queues an exit order for human approval. The engine surfaces the decision — it does not override your stop-loss policy.
Sometimes the hardest part is reaching out — but once you do, we'll make the rest easy.
Let's talk today
No-risk guarantee: if we can't find automation worth more than it costs to build, you owe us nothing — and you keep the roadmap.
Prefer email? info@chronexa.io
Automation Audit Request
We'll review your workflows and suggest where AI can save time & cost.