An order-flow assistant for CME Micro E-mini S&P 500 futures. It rebuilds the full order book from exchange order-by-order data, tracks large resting orders, runs a set of detection agents that estimate whether an institutional algo is active, and turns that context into trade signals I place by hand. Then I tried hard to prove it didn't work.
When an institution buys or sells size, it rarely does it in one print. It rests large orders in the book and works child orders through the market over minutes. A trader who can see that activity early can ride the move. Building the tool to see it turned out to be the easier half. The harder half was building a testing process strict enough that a good backtest actually means something, because a backtest will happily tell you whatever you tune it to say.
Reads CME market-by-order data from Databento, about 10 million add, cancel, modify and fill messages per session, and rebuilds the book with exchange queue rules. Output is a per-second depth snapshot plus an event log for every large resting order: placed, refilled, moved, partially filled, pulled.
A heatmap of resting size over time with trade prints overlaid, tracked large orders labeled by size and queue position, and a depth ladder showing each big order's share of its price level.
Four independent detectors look for an algo working an order: size clustering, a Hawkes-style rhythm model, a Bayesian score and a hidden Markov model. A logistic fusion step combines them, and a head collaborator explains in plain English when they agree and when they don't.
A rules-based layer combines market context with order-book events into entry and exit signals, with a hard stop, a trailing exit and a daily loss limit. It extends the decision logic from my Apex Overlay.
Replays real sessions event by event in the browser, or runs dozens of sessions in seconds. The same engine code runs in Node.js for batch testing, so a backtest and the dashboard always match.
A strategy tuned on the same data it's scored on will always look good. These rules existed to stop that from happening.
Settings were tuned only on September 1 to 15. All of August and September 17 to 30 stayed locked in the dashboard until the settings were frozen. When I accidentally saw one sealed day during testing, I excluded it from the results.
The dashboard logs every backtest, including re-runs, so it is visible how many versions were tried. Changes were chosen on the development days, then checked once on sealed data. The sealed data never guided a change.
Entries include a human reaction delay. Stops are modeled as resting broker orders that fill on the first touch, after testing showed a periodic-check model was hiding losses.
Yearly Sharpe and Sortino come from daily P&L including flat days, the industry convention, with per-trade values shown alongside. Fees are applied to every reported result.
| Sealed data, after fees | Core | Filtered |
|---|---|---|
| Trades | 37 | 15 |
| Win rate | 57% | 60% |
| Net P&L, 1 MES | +$69 | +$126 |
| Profit per trade | $1.87 | $8.42 |
| Yearly Sharpe | 1.4 | 3.8 |
| Yearly Sortino | 1.8 | 7.5 |
31 sealed sessions (August and September 17 to 30, 2026), one Micro contract, about $1 per trade in fees. Fifteen trades is a small sample, and ratios this high usually come down with more data. Backtested results only. Not financial advice.
Next: a live data connection for paper trading, then a final sealed test on October before any real money.
Every individual order the exchange receives, with its own ID: adds, cancels, modifications and fills. It shows each order in the queue, plus the total size at each price.
At each price, orders fill first in, first out. Knowing where a large order sits in the queue tells you how much has to trade before it starts filling.
Data held back until the settings are frozen, then scored once. It is the closest a backtest gets to trading the future.
Return divided by volatility. Sharpe counts all swings; Sortino counts only losing ones. A Sortino well above the Sharpe means losses stayed small next to the wins.
A futures contract worth $5 per index point, a tenth of the E-mini. Each 0.25-point tick is $1.25, which makes it a practical contract for testing a strategy with real money at small size.
Built as a product, run like an experiment: a clear hypothesis, a testing process designed to catch my own mistakes, and decisions driven by data I wasn't allowed to see in advance.