Company Watch
The pothole avoider. We take out the downsides so you can enjoy the upsides.
Do Nothing
-3.6%
Average loss across 5 stocks
Worst: BABA at -12.8%
Let NOAH Trade
-2.1%
Average gain over doing nothing
Best: NTES saved +8.5%
Average Alpha
-2.1%
The market fell. NOAH got you out. When you buy back at the bottom, that saved loss is your profit.
| Stock | Do nothing | NOAH saved you |
|---|
| BABA | -12.8% | -1.9% |
| JD | +11.5% | -11.2% |
| BIDU | -6.1% | -3.8% |
| NTES | -8.3% | +8.5% |
| PDD | -2.2% | -2.0% |
Special Operations
What if we shorted the potholes instead of just driving around them?
Right now NOAH sells before a drop and buys back after — the pothole avoider. But what if NOAH had also gone short during those drops? Every avoided loss becomes a direct profit, on top of the gains from buying back cheaper.
NTES — from -8.3% to
+8.7%
Avoided the loss + profited from the fall
| Stock | Do nothing | NOAH saved | If shorted too |
|---|
| BABA | -12.8% | -1.9% | -16.6% |
| JD | +11.5% | -11.2% | -11.0% |
| BIDU | -6.1% | -3.8% | -13.6% |
| NTES | -8.3% | +8.5% | +8.7% |
| PDD | -2.2% | -2.0% | -6.2% |
Hypothetical. Shorting carries additional costs (margin, borrow fees) and risks not modelled here.
Full transparency. This is a live experiment, not a fund. We publish every trade, every loss, every alpha calculation. We are testing whether concentrated AI intelligence — reading daily reports, running autonomous due diligence, and making real-time decisions — can outperform the simplest strategy of all: buy and hold.
The idea comes from the Berkshire Hathaway school of investing: focus on one company, know its environment deeply, and make decisions from a position of genuine understanding. We took that principle and gave it to an AI.
For each stock we track, we run two parallel portfolios side by side:
- The Active line — An AI reads daily intelligence reports, runs its own due diligence, and decides whether to buy, sell, hold, or duck. It has stop-losses, profit-taking rules, and can override report advice if it spots danger.
- The Passive line — Buys on day one and never sells. The simplest possible strategy. The benchmark to beat.
Alpha is the difference between them. Positive alpha means the AI is winning. Negative means we’d have been better off doing nothing.
How It Works
The system runs a continuous daily cycle across 5 stocks, polling intelligence reports via RSS and making autonomous trading decisions.
1. Report
Daily intelligence report arrives via RSS at 1pm UK time. Contains stance (BUY/SELL/HOLD/FADE) and confidence score for each stock.
2. Assess
AI forms its own “house confidence” independently of the report. Checks news wire, Hong Kong tape, and macro. Dual-confidence system prevents blind trust.
3. Decide
BUY or SELL at 65%+ confidence. HOLD at 45–64%. FADE below 45%. No emotion, no FOMO — mechanical thresholds only.
4. Defend
Autonomous due diligence every 2 hours. Stop-losses, profit-taking, drawdown protection, market crash override. Multiple layers of defence.
5. Measure
Compare active P&L against passive buy-and-hold. Alpha is the verdict. If the AI can’t beat doing nothing, we want to know.
The Defences
The active line isn’t just making buy/sell decisions. It has a layered defence system designed to protect against the inevitable bad days that destroy buy-and-hold returns.
Duck-and-Cover
Pre-market DD checks the Hong Kong tape and overnight news before NYSE opens. If a storm is coming, the system sells at 9:30am and re-buys at 10:30am — only if the thesis survives. Designed to avoid the first-hour bloodbath.
Autonomous Due Diligence
Every 2 hours, the AI independently checks if the investment thesis still holds. It monitors the news wire, checks the Hong Kong tape, and watches SPY and VIX for macro shifts. The “human in the loop” that never sleeps.
Profit-Taking
At +15%% the system considers taking profit. At +25%% it strongly considers it. If the stock pulls back 5%% from its peak, gains are protected automatically — drawdown from peak triggers an exit regardless of thesis.
Stop-Loss (Soft + Hard)
At −10%% the AI assesses whether to cut — asking the LLM if the thesis is still intact. At −15%% it exits immediately, no questions asked. The hard stop is mechanical and absolute.
Market Crash Override
If the S&P 500 drops more than 3%% in a session, all long positions are exited immediately. Individual stock thesis doesn’t matter when the whole market is crashing — everything correlates on the way down.
Horse-Bolted Detection
If the stock has moved more than 8%% since the last report, the system flags a “horse bolted” scenario — the move the report anticipated may have already happened, and entering now is chasing rather than front-running.
Live Portfolio
Last updated: 2026-04-27T06:45
BABA
Alibaba
$135.82
FLAT
Stance: HOLD | Confidence: 55%%
35 trades | 67 days tracked
JD
JD.com
$30.27
FLAT
Stance: HOLD | Confidence: 55%%
35 trades | 67 days tracked
BIDU
Baidu
$128.71
FLAT
Stance: FADE | Confidence: 55%%
4 trades | 67 days tracked
NTES
NetEase
$110.58
FLAT
Stance: FADE | Confidence: 55%%
32 trades | 67 days tracked
PDD
PDD Holdings
$98.06
FLAT
Stance: HOLD | Confidence: 55%%
2 trades | 67 days tracked
Honest Performance
The alpha paradox. Our best-performing stocks by alpha are the ones the AI has never traded. By staying flat while the stock fell, the AI avoided losses that the passive line absorbed. The AI is winning — but it’s winning by not playing, not by playing well. We are being transparent about this because it matters.
-2.1%%
Avg Alpha (Active vs Passive)
Of 5 stocks tracked, the AI is beating passive on 1. Best alpha: NTES at +8.5%%. Worst: JD at -11.2%%. The system is still in its early learning phase — we need more data points and more market conditions before drawing conclusions.
The Trading Cost Question
A fundamental question hangs over the entire experiment: is buying and selling shares the best instrument for this strategy? Our active line has generated 108 trades across 5 stocks in 67 days. In a real-money implementation, every one of those trades would carry costs.
These are the instruments we’re investigating:
Individual Shares
Current approach. Direct exposure, maximum flexibility. But commission costs, bid-ask spreads, and potential pattern day trader rules could erode returns on high-frequency active trading.
ETFs (e.g. KWEB, FXI, MCHI)
China-focused ETFs offer basket exposure with lower per-trade impact. Reduces single-stock risk but dilutes the “know one company deeply” principle. Lower spreads on liquid ETFs might make the duck-and-cover strategy more cost-efficient.
Options
Protective puts could replicate duck-and-cover without selling the underlying. Covered calls could generate income on flat periods. But premium decay (theta) and complexity introduce new risks. Options require a different confidence threshold.
CFDs / Spread Bets
Leveraged instruments with no stamp duty (UK). Lower capital requirements for the same exposure. But overnight funding costs accumulate on held positions, and leverage amplifies the very losses our defences are designed to prevent.
We will be modelling the cost impact of each instrument against our actual trading history to determine which vehicle best suits a high-frequency ducking strategy. Results will be published here as we learn.
What We’re Learning
This system is live and evolving. Here are the key findings from our first 67 days:
Not trading is a strategy
Our highest-alpha stocks are the ones the AI stayed flat on while the passive line lost money. The AI’s best move has been inaction — correctly identifying when NOT to enter. This is valuable but raises the question: is the alpha coming from intelligence or from caution?
Win rate needs work
At 3.5%% win rate across 108 closed trades, the majority of individual trades lose money. The system may be churning — entering and exiting too frequently without conviction. We are investigating whether wider entry thresholds and longer hold periods would improve outcomes.
Early days — we need more data
We have not yet seen a market crash, an earnings surprise, a geopolitical shock, or a sustained rally. All conclusions are preliminary. We need months, not days, to assess whether the defensive mechanisms actually add value across different market regimes.
The dual-confidence system works
The gap between report confidence and house confidence is proving useful. When the report says BUY at 40%% but the AI independently rates it at 55%%, the system correctly holds rather than entering. This separation prevents blind-following and has contributed to the positive alpha.
Why publish this? If AI-managed active trading can consistently beat buy-and-hold, that changes how individual investors think about stock ownership. If it can’t, we want to know that too — and we want anyone following our work to see the honest evidence. This is not a product. This is a live experiment, and you are watching it unfold.