Case studies

Case 1 — COIN live trade (winner + exit)

One live brokerage trade: buy on a high G-score, sell on a Model Signal Framework downgrade.

This is one live trade from my own brokerage account. It is for transparency so you get a sense of how the model makes decisions.

I am showing a winner because the useful part is the exit, not the entry. The exit signal uses the same ticker and the same scoring framework, but makes a different decision once the score changed.

Live brokerage proof

The setup (Feb 12, 2026)

On 2026-02-12, COIN was a top-3 signal in my stack. I bought at $140.20 in my live account.

The model snapshot from that day showed:

• G-score: 101.81%
• Price: $141.09

That was a clean “information still matters” read: COIN’s target distribution sat well above market price. From the screenshot below, analyst sentiment still carried some dispersion ($54.10 standard deviation), but was centered far above where the stock was trading.

The exit (Mar 16, 2026)

On 2026-03-16, the G-score dropped below 10%. In my rules, that is a Model Signal Framework exit signal since it fell below the live 65th percentile cutoff at the time. I sold at $204.80.

The model snapshot from that day showed:

• G-score: 9.89%
• Price: $203.32

Same instrument and same measurement rules, but with a different output → a different action. From the screenshot below, the analyst sentiment distribution had shifted enough that COIN was no longer mispriced in the model’s eyes.

This is the part other models skip because it’s boring: the exit was not a call on crypto (in the case of Coinbase) or macro. It was a mechanical downgrade on the ticker using the same measurement method as the entry.

Returns

• Entry: $140.20
• Exit: $204.80
• Holding period: 32 calendar days (2026-02-12 to 2026-03-16)
• Absolute return: +46.08%
• Annualized equivalent (math extrapolation only, not a forecast): about 7,438%

The trade itself was live in my account, with real slippage, fees, and tax effects not fully detailed here. Past performance, including this single-name example, does not indicate future results. This is general publication content from The Alpha Engine Report, not personalized investment advice. n=1 case study.

That annualized figure is exactly why short-term holds need a footnote. It answers the math question, but it should not be treated as expected CAGR.

Opportunity cost check after exit

Since the sale, COIN has been mostly flat/slightly down versus my exit print.

I redeployed capital into another name with a higher score under the same G framework. This case study is about process, so I am not naming that ticker here.

For G-score construction and execution logic, read the methodology page. Please also read disclosures alongside any performance figures.

Case 2 — BIRK legacy exit (frozen stop + redeploy)

One live sequence: legacy position, frozen coverage, systematic exit, capital redeployed into a higher-scoring signal.

This is transparency from my own account to show a more complex example of the model dynamics. I am showing a series of executions that didn’t always feel good. One chart will make the exit look early. The point is process discipline under one set of rules.

Model Signal Framework for existing positions

An important note before we get into the case: the model’s execution framework is not just for new entries. It is also for legacy holdings that were opened before a formal scoring process. If old positions sit outside your measurement rules, you are unnecessarily running two portfolios in your head.

The setup: a large legacy BIRK position

I carried a large losing BIRK position from last year, before I built the Alpha Engine. On 2026-03-20, the model’s output flagged BIRK as Frozen (i.e. insufficient data) with a price of about $33.34. That meant the analyst-target structure was too thin for a trustworthy G-score at that snapshot. At that point, it stopped being “a name I like” and became “an unreliable dataset.”

The execution: frozen stop + 25% fade + remainder (2026-03-27 to 2026-03-30)

In the model logic, a frozen state waits to action until coverage improves or a frozen-stop triggers.

In this case, on 2026-03-27, the frozen-stop condition fired. I sold the first tranche at $34.32. However, I did not force a one-print full exit. I used the model’s 25% fade size limit (yes, I had that big of a position) because size and execution quality matter.

On 2026-03-30, I sold the remainder at a better price of $34.61. The fade saved me nearly a percent.

The skeptic’s chart objection

The chart above shows a clear question about the model’s exit timing: BIRK rebounded after my exit.

As of the time of drafting (May 25), BIRK has been quite volatile, but mostly higher than my exit, with an average price of $37.56 since March 30, roughly +9% versus my first exit print at $34.32. Its recent buyback news even lifted its price back into the $40s.

That is exactly why this case matters. The question is not “did I call the local top.” The question is “did I follow the same measurement system when coverage broke” and whether the freed capital was better used elsewhere.

Opportunity cost: what the BIRK exit funded

Exiting BIRK released risk budget for a higher-ranked signal in the same framework: on 2026-03-27 and 2026-03-30 (same days I closed BIRK), I redeployed this capital into RGTI at a blended average cost of about $13.23.

The 2026-03-27 model snapshot for RGTI showed a G-score of 82.39% at the time:

I then closed RGTI on 2026-04-23 at $17.29, for about +30.7% on that round trip.

The trade itself was live in my account, with real slippage, fees, and tax effects not fully detailed here. Past performance, including this single-name example, does not indicate future results. This is general publication content from The Alpha Engine Report, not personalized investment advice. n=1 to n=2 case study. Opportunity-cost comparisons depend on position sizing and timing specific to my account.

In my account-level dollars, this redeploy path vastly outperformed simply holding BIRK through the same window. Looking at the “bad exit” in BIRK without also looking at the model’s subsequent decision gave an incomplete view of the framework’s decisions.

After exiting RGTI, I rolled proceeds into the next ranked setup on the same playbook. The same bookkeeping lesson holds: when information quality breaks, release capital cleanly and reallocate under one scoreboard.

For G-score construction and execution logic, read the methodology page. Please also read disclosures alongside any performance figures.

Case 3 — COVID stress (Mar–Jun 2020 backtest)

One stress-window case study from backtest logs: drawdown, dip refills, and Q2 follow-through.

This is one stress-window case study from my backtest logs. It is about what the rules did when marks got bad, and how the model opportunistically added additional risk to outperform the market during a correction.

The setup (March 2020)

On 2020-03-02, the total portfolio value was $2,144.80. By 2020-03-19, the unrealized PnL was -$1,992.

As expected during a market downturn, in that stretch the system fired 9 dip refills totaling $2,436.20.

What happened at the ticker level

March was not a passive hold but was full of dip refills. The action log showed methodical new buys, with the following top 5 tickers bought during this market downturn: FANG, SNAP, LYFT, UBER, and APA.

Q2 follow-through (through 2020-06-30)

By 2020-06-30, total portfolio value reached $9,951.01. SPY under the same flow schedule was $7,311.05 at that point (about a 1.36× gap in portfolio value). Note that for parity, the index portfolios also executed dip refills when the market plummeted in March.

In the end, each of the five names above beat SPY and QQQ over the Mar to Jun window. Calculating cash flow using the backtest logs gets us an XIRR of 76.5% for 2020-03-02 to 2020-06-30, far exceeding that of SPY and QQQ.

All figures in this case study are from a hypothetical backtest. They do not reflect live trading results and do not guarantee future performance. Regime-specific outcomes from 2020 may not repeat. See disclosures for definitions and limitations.

For G-score construction, dip-refill logic, and execution rules, read the methodology page.