Trading the emotions behind corporate scandals
Property mode (beta): where the housing cycle is, how prices are moving by country, city and region, and what a purchase would cost you.
Heat map
Pinch or scroll to zoom, drag to move, tap a circle for details. Red is falling and green is rising; a bigger circle means a bigger move. Grey means no published figure yet. Real state and province boundaries; tap a region for its name.
By area
Colour shows the latest change in home values: red is falling, green is rising, grey means no published figure yet. Point at a tile for the details and source. Refreshed weekly.
Market signals
| Signal | Latest | As of | Direction | What it means |
|---|
What would it cost me?
My shortlist
Add homes you’re considering. Each one is costed with the deposit, rate and term set above, checked against the latest trend for its area, and saved in this browser only.
Ready to buy? Get connected
A purchase can’t be completed inside an app: it needs a lender, a contract and a conveyancer. These are the people who make it happen, and where to find accredited ones.
How a purchase works
- Get pre-approval through a broker or lender, so you know your limit and can move fast.
- Shortlist and inspect. Use the shortlist above to compare costs and the heat map to see where values are heading.
- Do your checks: building and pest inspection, strata report for units, and a contract review by your conveyancer.
- Make an offer or bid at auction. Private sales usually have a cooling-off period; auction sales don’t, so finish your checks first.
- Exchange contracts and pay the deposit (often 10%, sometimes less by agreement).
- Settle: your lender pays the balance, stamp duty is paid, and you get the keys.
Suburbs and neighbourhoods
Published suburb-level figures are patchy and mostly sit with paid data providers. Tell me the suburbs you’re watching and the weekly update will add each one to the heat map whenever new figures are published. Brisbane, Perth and Adelaide are falling almost everywhere: virtually every suburb in those cities is down over the past three months.
General information only, not financial advice. Stamp duty and grants vary by state: check your state revenue office’s calculator.
This robo-trader reads company scandal news and bets on how people react that day. Right now it practises with pretend money and real prices.
Live practice account
Day-trading with pretend money since 18 Nov 2026. Every hour the US market is open it scans the news, trades new stories, and closes everything by the final bell. Updated —.
| Date | Stock | Story | Call | Stock moved | Result |
|---|
Is it working?
—
What kind of news it’s good at
| Type of story | Examples | How it does | Right |
|---|
Its latest calls
Each call is made at the opening bell and closed at the end of the same day, on a $100,000 practice account.
| Date | Company | What happened | Robo-trader | Result |
|---|
Try it yourself
How it decides
What it watches
Any listed company, any size. New kinds of news are watched and learned from first, and traded once it has seen 8 examples of that kind.
- Company scandals and shocks: CEO exits, misconduct, fraud and short-seller reports, lawsuits, breaches, crashes. Trading now.
- Drugs, biotech and medical devices: FDA approvals, rejections and device clearances, trial results, recalls. Learning.
- Healthcare: insurers, hospitals, Medicare and drug-pricing decisions. Learning.
- Tech: AI, chips, cloud and software: big launches, outages, export rules, antitrust. Learning.
- Crypto: bitcoin, ether, exchange hacks, regulation; traded through bitcoin and ether funds and crypto stocks. Learning.
- Cannabis: rescheduling, banking rules, legalisation and company news. Learning.
- Berkshire Hathaway: quarterly holdings filings and stake changes. Learning.
- Politics: tariffs, sanctions, elections, wars. Learning.
- Commodities: oil, gas, gold, copper, grains, through funds. Learning.
- Politicians’ stock trades: disclosed buys and sells by members of Congress. Learning.
- Coming later: hedge-fund mode: holding for days, hedging with sector funds, pairs trades and investor reports.
Words you might see
- Bet against (short): a trade that makes money if the price falls.
- Opening gap: how far the price jumps or drops at the opening bell compared with the day before.
- Practice (paper) trading: pretend money, real prices, real timing.
- Never seen: scandals added after the model was locked, so it could not have been tuned to them.
Eleven human emotional correlates from the behavioral-finance and affective-science literature, turned into daily indices, weighted by a relevance metric and mapped onto NVDA trading before, during and after the news: the Oct 29 spy-link report, Huang’s Nov 8 admission, his Nov 12 resignation, and the Nov 17, 2026 earnings under interim CEO Colette Kress.
The Huang case in one view
NVDA moved — from the Nov 2 close to the post-earnings after-hours print. Of the emotions measured, — carry the most relevance. The relevance-weighted composite explains — of the variance in daily abnormal returns in-sample, but out of sample (leave-one-out) its R² is —: on twelve sessions it maps the episode well but does not yet forecast.The relevance-weighted composite explains — of the variance in daily abnormal returns in-sample and holds up out of sample (leave-one-out R² —).
Price and abnormal return, session by session
Abnormal return = NVDA return − 1.5 × Nasdaq return (beta assumed). Bands mark the three phases. Point at a session for its news.
Which emotions matter: the relevance metric
R = prior1−λ × fitλ with λ = 0.35. The prior scores each emotion’s evidence base, its specificity to CEO-misconduct and departure events, and how observable it is here; fit is its correlation with abnormal returns in the direction theory predicts (zero when it points the wrong way). The dark tick marks the prior alone.
Trade ticket
Describe a news event on a heavily traded stock. The ticket gives the model’s call for the open-to-close leg, with its probability, expected move and position size. Every number comes from the fitted model on the Accuracy tab. Paper-trade only.
How the probability moves with the gap
Scandal news only. The shaded band is where the edge is too small to trade, and the dot marks the gap set above. Beyond the worst gap in the data (about −13%), the curve is an extrapolation.
Model v3.1: the seven improvements
Across — scandals, the frozen persona model’s predictions correlate — with the actual open-to-close moves, against — for the old gap-reversal line. Its scandal-level p-value is —: closer to 0.05, but not there.
| # | Improvement | Status | What was done |
|---|---|---|---|
| 1 | More scandals | Partial | 15 scandals, up from 7. Added Wynn, Lululemon, Intel, Papa John’s, Texas Instruments, Tesla 2018, Boeing 2019 and Carvana 2025. Guess, Block, Icahn, Super Micro, Activision and WWE need a data feed. |
| 2 | Persona-specific signals | Partial | Bargain-hunter attention now comes from the event-day volume surge. Options, fund flows and short interest are waiting on a connector. |
| 3 | Intraday bars | Blocked | Needs a market-data connector (logged below). |
| 4 | More human personas | Done | Habituation (each earlier shock in the same story softens the reaction), bigger responses allowed at extremes, and resolution and threat read from the news itself. Cost-basis anchors are waiting on ownership data. |
| 5 | Automated news scoring | Done | A fixed word list scores threat, resolution and betrayal in each session’s news text. Valuation and the founder / under-performer facts are still set by hand. A second, blind scorer is still to come. |
| 6 | Proper testing | Done | The settings were frozen and fingerprinted before scoring. Tests run by scandal, on a clean out-of-sample set of 8 events, and the backtest includes slippage, spreads, borrow costs and a 25% cap. |
| 7 | Learning | Done | Persona weights update one scandal at a time (predict, then learn), held close to the research priors. |
Model comparison
| Model | Correlation | RMSE | Big-gap direction | Correlation, fresh | Direction, fresh | Scandal-level p |
|---|
Predicted against actual, every scandal-news session
What the 15 scandals teach
- Unresolved legal threats keep falling. Tesla’s SEC suit and Wynn’s open investigation both kept dropping after big gaps down. The personas called this; the old fade-the-gap rule bet the wrong way.
- Resolution lets buyers back in. Nikola’s founder quitting and Huang’s resignation both reversed hard upward.
- Relief rallies keep going. Papa John’s, Kohl’s and Boeing kept rising after their CEOs were pushed out, but the personas wrongly fade those gaps. This is the next idea to test: freeze it as v3.2 and score it only on new scandals, not by retuning on these 15.
- The trade-level p overstates the case. The backtest’s sign-flip p of about 0.02 counts five Nvidia days as separate trades. The scandal-level p is the one to trust.
Realistic backtest, post-Bezos
| Date | Session | Gap | Predicted | Actual | Position | P&L | Account |
|---|
Frozen specification and learned weights
| Setting | Value |
|---|
| After scandal | Worrier | Bargain | Value | Risk | Follower | Contrarian |
|---|
Connections log: to add later
Logged 18 Nov 2026. Each connection unlocks a blocked or partial improvement.
| Connection | Unlocks | Improvement | Status |
|---|---|---|---|
| Alpha Vantage | 1–5 minute intraday bars, options chains and implied volatility, news sentiment | 2, 3, 5 | To add |
| Financial Datasets | Daily prices for delisted and missing names (Activision, WWE, Nikola day 1, Guess, Block, Icahn, Super Micro), ownership and insider data for cost-basis anchors | 1, 4 | To add |
| Clear Street | Options activity and instrument events, a short-interest and borrow-cost proxy for the Contrarian | 2, 6 | To add |
| MT Newswires | Time-stamped headlines for automated scoring and the live scandal monitor | 5, monitor | To add |
| Questrade | Brokerage for paper, then live orders, only after the forward test passes | Auto-trading | To add later |
| Scheduled tasks permission | Daily post-close scandal scan with push alerts and paper positions | Monitor | Blocked by setting |
The people behind the price
On a scandal morning, six kinds of people trade the stock, and each asks a different question. Their combined buying and selling predicts the move from the open to the close. Each persona’s share of trading and its decision rule were set in advance from research and market structure, not fitted to the outcomes.
How well the personas predict
Who moved the stock, session by session
Bars split each predicted move by persona: right means buying, left means selling. The dot is what actually happened. Point at a row to see its descriptors.
Reading the people
- Resolution is what lets people buy. When Huang resigned and an interim CEO was named, the Worrier and the Risk Manager had nothing left to fear. Bargain hunters, value buyers and contrarians then had the stock to themselves, which drove the biggest rebound in the data.
- Open threats keep sellers in. On Nov 9, with an FBI inquiry and a board probe still open, risk managers and trend followers offset the dip buyers. The model expected a small bounce, and the real one was modest.
- People misread relief. At Kohl’s, the value investors saw the jump as overdone. Real holders treated firing an unpopular CEO as better news than the model allowed.
- Small gaps leave everyone on the sidelines. Below about 3%, no persona has a strong reason to act, which is why small scandal gaps don’t reverse reliably.
- The model gets direction better than size. It caught the direction of the big rebounds but only about half of their size. Panic, attention and bargain hunting are probably non-linear at the extremes.
Making the personas more human
- Give each persona its own data. Retail order flow for bargain hunters; put/call ratios and implied volatility for worriers; ETF and fund flows for risk managers; short interest and borrow cost for contrarians.
- Time of day. Worriers sell in the first 30 minutes, value buyers come after the panic, and risk desks rebalance into the close. Intraday bars would let each persona act on its own clock.
- Habituation. Each repeated shock in the same story moves people less, as on Nov 10 and Nov 16. Scale reactions by how many shocks came before.
- Different anchors. Loss aversion depends on what each person paid. Retail often bought near the highs; long-term funds hold far lower. Use cost-basis estimates per persona.
- Trust repair. How fast and how credibly the board acts, and how complete the apology is, change how betrayed holders feel. Astronomer resolved in three days; Nvidia took four.
- Social contagion. Virality (search trends, social mentions) brings in new Worriers and Bargain Hunters. The Byron clip is the extreme case.
- Let the personas learn. Once 50+ scandals are logged, update each persona’s share and sensitivity, one scandal at a time, with shrinkage toward the research priors.
Descriptors used for each session
| Session | Gap | Threat | Resolved | Attention | Valuation cushion | Business damage % | Predicted | Actual |
|---|
Post-Bezos backtest: $100,000 in real time
This is the model as it would have run live after the Bezos scandal. It was first fitted only on the 2019 Amazon sessions. Before each later scandal it was refitted on what had happened so far, bought or shorted at the open, and exited at the close. Known bias: the scandals were picked with hindsight, and the fade-the-gap idea came from Nvidia, which is also in the test.
$100,000 became — (—) over — trades. Without the Nvidia sessions it would have ended at —; without Nov 12 alone, at —.
Account value, trade by trade
Every trade
| Date | Company | What happened | Gap | P(up) | Call | Open→close | P&L | Account |
|---|
Against simple rules and hindsight
| Strategy | Ends at | Return | Trades | Hit | Brier | Worst drawdown |
|---|
Is it statistically significant?
One hypothesis, fixed before testing: on scandal days, the intraday move goes against the opening gap. Every test is one-sided, uses 20,000 random reshuffles, and is Holm-corrected for the number of tests shown. The magnitude test clears p < 0.05 after correction, but it treats all 30 sessions as independent and leans on Nikola and Nvidia’s biggest gaps. The rank test, the out-of-sample events and the honest test (one value per scandal) all fail. The evidence is suggestive, not significant: more big-gap scandals are needed, not more tuning.
| Sample | Test | n | Statistic | p | p (Holm) | Verdict |
|---|
Reshuffled outcomes
Final account value when the 10 real results are drawn again at random, 10,000 times. The line marks the $100,000 start.
Scandals left out, and why
| Date | Company | Event | Reason |
|---|
Two opening prices are estimates: Nikola (from a reported ~25% premarket drop) and Nestlé (an indicated −1.9% open). Nikola is the second-biggest winning trade.
Trade simulator
Run the micro-trade on simulated scandal sessions. Each run draws opening gaps from real scandal sessions, and a reversal strength from the model’s own uncertainty. Change any setting and the results update. These are simulations under the model’s assumptions, not forecasts.
Equity paths
Bands hold the middle 50% and 90% of runs; the bold line is the median run and the thin line one sample run.
Your settings on the real sessions
The same rules applied to the 25 real open-to-close legs (NVDA 2026, AMZN 2019), using probabilities from models that never saw each leg.
Sensitivity grid: 2,000 runs per setting, computed in Python
Reversal 0 means the effect is not real: under that assumption every setting loses money in most runs.
| Gap ≥ | Kelly | Reversal | Cost bp | Median run | 5th pct | P(loss) | Median drawdown | Trades/run |
|---|
Emotion intensity by session
Each index runs 0–1, built from coded news intensity and the market’s own signals (opening gap, intraday drawdown, range, volume surge, distance from the peak). Rows are sorted by relevance.
The composite against what happened
Composite E = Σ R·sign·I ÷ Σ R, then abnormal return = a + b·E. Nov 5, a sector-wide rout with no benchmark, is left out of the fit.
Trading the emotions
Transient emotions (fear, panic, attention, hope, anger, herding) are faded, since search-based fear, media pessimism and disaster studies find they reverse within days. Persistent ones (betrayal, grief, ambiguity, loss anchoring) are followed, since managerial-indiscretion studies find losses that keep building for a year. Position = tanh(3 × (gate × contrarian + drift)), set at each close and held to the next print. In-sample, no trading costs.
Note how the long half-life of betrayal keeps the model short into the earnings print, which no emotion index could see coming: the main thing to fix with options and order-flow data.
Before, during and after
Pre = Nov 2–6 (spy-link reporting, no admission; Nov 5’s sector-wide rout excluded). During = Nov 9–12 (admission, probe, second affair, resignation). Post = Nov 13–17 incl. after-hours earnings. Emotion columns are mean indices.
| Phase | Return | Σ abnormal | Fear | Betrayal | Panic | Ambiguity | Grief | Hope |
|---|
The emotion taxonomy, stratified
Sign is the expected same-day effect on returns; half-life is in trading sessions (state = recomputed daily, no decay).
| Emotional correlate | Tier | R | Prior | Fit | Sign | Half-life | Key literature |
|---|
Retro-test: the Jeff Bezos scandal (AMZN, Jan–Feb 2019)
The same eleven variables, scales, half-lives and priors, run on Amazon: the Jan 9 divorce announcement, the Jan 10 National Enquirer affair exposé, the Jan 13 ‘Bozo’ tweet, and Bezos’s Feb 7 ‘No thank you, Mr. Pecker’ extortion post. The Feb 1 earnings drop is kept as a non-scandal control and left out of every fit. The transfer test applies the NVDA-fitted weights and slope to AMZN without refitting, so it is genuinely out of sample.
The same emotions rank as most relevant in both scandals (rank correlation —), and their mix during the news is alike (profile correlation —). What differs is intensity: no founder exit and no national-security fear, so AMZN’s event-window abnormal return was — against NVDA’s —.
Relevance of each emotion, Huang vs Bezos
How strongly each emotion ran during the news
Mean index over the ‘during’ sessions (0–1). Same shape, roughly half the amplitude, and grief, the founder-loss emotion, absent for Amazon.
Out-of-sample: the NVDA model predicting AMZN
The model gets direction and timing right but overshoots size on Jan 10: weights learned on a spy-linked founder exit overstate a divorce-and-tabloid story. Intensity calibration across scandal types is the next thing to learn.
Version 2: corrected model, people layer and micro-trades
Trading only the open-to-close leg, with probabilities validated on sessions the model never saw, the micro-trade made — across — trades, against — for holding every leg. Shuffling returns against signals beats it only — of the time, but the 90% bootstrap range of its summed P&L, —, still reaches into losses, and almost all of the profit came from two NVDA panic opens.
Accuracy scorecard
— of — tests passed. Verdict: —. The strategy makes money in tests that include the big NVDA panic days. It fails the stricter ones: probability accuracy, the bootstrap worst case, and training on Nvidia then testing on Amazon.
| Test | Result | Pass bar | Outcome | How it was run |
|---|
Are the probabilities honest?
How often a leg actually closed above its open, compared with the probability the model gave it, out of sample. Honest probabilities make the two bars match. With 25 legs every bin is small.
What was corrected
| # | Correction | Failure it fixes | Change |
|---|
Whose emotions: traders, shareholders, stakeholders
The same event is appraised differently by each group, through its own goals and horizon. Traders set the intraday price; long-term holders set the weeks-long drift; stakeholders (employees, regulators, customers, the board) are predicted rather than traded, and pass their emotions on through the next day’s news. Retail demand runs against returns: retail investors buy falling, attention-grabbing stocks and supply the liquidity others sell into.
| Actor | Group | Horizon | Emotions (sensitivity, demand sign) | Corr. with abnormal return | Basis |
|---|
Actor emotion intensity, NVDA 2026
Actor emotion intensity, AMZN 2019
Other executive scandals: severity calibration
Day-0 moves reported in the news, grouped by class. Class means set the severity multipliers. Astronomer (Andy Byron, Coldplay kiss-cam) is private, so it has no price. It shows the stakeholder channel instead: viral attention led to leave, then resignation, within three days.
| Company | Date | Executive | Class | What happened | Day 0 | Worst intraday | Recovery off low | Source |
|---|
Micro-trades, out of sample
One bar per open-to-close leg. Each P(up) comes from a model fitted without that session, sized at a quarter of the Kelly fraction (a standard bet-sizing rule), net of 5 bp costs. The overnight leg showed no skill out of sample, so it stays flat, and it is always flat into earnings.
Simulations
Left: P&L when returns are shuffled against signals 2,000 times; the line is the real result. Right: open-to-close returns for 20,000 simulated scandal sessions with a gap of −5% or worse. Gaps are drawn from the cross-section above; the reversal slope comes from its bootstrap.
Beyond scandals: any news on high-volume stocks
Trigger: a top-50 most-active stock trading at 2× its 30-day volume or gapping 3% or more. The news type decides whether to fade the gap or follow it. In the sessions so far, emotional shocks reversed and information shocks (earnings, a sector rout) carried on.
| News type | Micro-trade rule | Dominant emotions | Sessions in the model | Gaps that reversed | Basis |
|---|
Forward-test plan before any live trade
- Freeze the model. Lock today’s settings (gap-only probability, quarter-Kelly, 5 bp costs) and do no refitting during the test.
- Log before the open. For every trigger (a top-50 most-active stock at 2× volume or a 3% gap), record the news type, class, gap, P(up) and paper position before 9:30 ET.
- Score at the close. Open-to-close return and P&L after costs, Brier score and hit rate, plus 1-, 5- and 20-day drift to test the slow emotions.
- Widen the history. Rebuild 30+ past executive scandals with daily open/close data, and have a second, blind coder score the news so the scoring can be checked.
- Set the pass bar in advance. After 30 triggers, require a Brier score below 0.24, direction hit above 55%, positive P&L after costs and a permutation p-value under 0.05. Only then consider small live orders through a connected broker.
Model selection grid
| Features | Ridge | Brier | Hit | Cum. P&L % | Trades |
|---|
The richer actor features fit the past better but gave worse probabilities on unseen sessions, so the traded model keeps only the opening gap. The actor layer stays as an explanation of who is moving the price.
Model and code review: gaps, ranked
An audit of this build, most serious first. Severity: Critical can make the headline results wrong; High overstates confidence; Medium limits realism or makes the code fragile.
| # | Severity | Gap | Why it matters | Fix |
|---|---|---|---|---|
| 1 | Critical | Too few independent scandals | Seven scandals, and only five sessions with gaps of 3% or more. Treating nine Nvidia days as independent turns p ≈ 0.15 into p ≈ 0.005. | Build a 60–100 scandal event set with exact daily open/close prices; test by scandal, not by day. |
| 2 | Critical | Hindsight in design and coding | The fade rule was inspired by NVDA, which is in every test, and the news scores were hand-coded by one coder who knew the outcomes. | Pre-register v2.2 and freeze it; score news automatically (time-stamped headlines plus an emotion lexicon) with a blind second coder. |
| 3 | High | Selection not inside validation | Ridge strength and features were picked by the same leave-one-out runs used to report accuracy, and the permutation test does not redo that choice, so its p is optimistic. | Nested cross-validation; redo the selection inside each permutation. |
| 4 | High | Estimated or recalled inputs | Nikola and Nestlé opens are estimates (Nikola is the second-best trade), some 2019 Nasdaq moves were recalled, and Kohl’s day 0 is a range. | Load exact prices from a market-data connector or CSV; drop any event without them. |
| 5 | High | Circular severity calibration | Class multipliers came from a cross-section that includes NVDA and AMZN, then were applied back to them. | Calibrate severity on events outside the test set only. |
| 6 | Medium | Execution is idealised | Fills exactly at the open, a flat 5 bp cost, no trading halts (Kohl’s was halted), no short-borrow fees (Nikola was hard to borrow), and up to 100% of equity in one stock. | Model opening-auction slippage, spreads and halts; add borrow costs; cap any one stock at 20–25%; hedge with a sector ETF. |
| 7 | Medium | Probabilities not calibrated | Brier scores of 0.24–0.26 are coin-flip level, so the Kelly sizing rests on edges the model cannot yet measure. | Size from a fixed fraction until calibration passes; apply isotonic or Platt calibration once there are 50+ events. |
| 8 | Medium | Daily bars only | There is no intraday path, so entries after the first 30 minutes, stops, or VWAP exits can’t be tested, and the simulator treats worst-intraday moves as opening gaps. | Add 1–5 minute bars for event days. |
| 9 | Medium | Unused or untested layers | The people layer and the slow-drift emotions (betrayal, grief) explain moves but aren’t validated for trading; v1 relevance weights are fitted in-sample. | Test the 20-day drift as a separate strategy; drop or shrink anything that doesn’t pass out of sample. |
| 10 | Medium | Fragile code structure | Episodes swap shared module settings at run time, features are read by position, the sizing logic lives in both Python and JavaScript, plus dead code and unused imports. | Use explicit config objects and named features; export one sizing spec read by both; add unit tests and data-validation checks. |
| 11 | Medium | No live signal record | Every result so far is a reconstruction. | Paper-trade journal saved on this page, with scheduled news monitoring once permissions allow. |
The algorithm
- Code the news. Each session’s public record is scored 0–3 per emotion.
- Read the market’s own affect. Opening gap (panic), intraday drawdown (fear and stress), recovery off the low (capitulation exhaustion), high–low range (disagreement), volume ÷ 30-day average (attention), distance below the 52-week high (loss-aversion anchor).
- Build indices. Ik,t ∈ [0,1] blends the coded and market parts on fixed scales, so different scandals sit on one ruler.
- Let emotions decay. Sk,t = Sk,t−1·2−1/hk + Ik,t, with half-lives from the literature: panic 1, fear 2, attention 2, hope and anger 3, ambiguity 5, grief 20, betrayal 60 sessions.
- Score relevance. Rk = prior1−λ × fitλ. Tier 1 at R ≥ 0.55, Tier 2 at R ≥ 0.35.
- Combine. Et = Σ Rk·signk·Ik,t ÷ Σ Rk, regressed on abnormal returns.
- Trade. position = tanh(3 × (gate × contrarian + drift)).
Data notes and limits
- Prices, OHLC and volume as reported by Gotrade (Yahoo Finance and Nasdaq data), Reuters, AP, Motley Fool and Investing.com.
- No options, implied-volatility, order-flow, search-volume or news-tone data were available here.
- Twelve sessions are far too few to estimate eleven weights: relevance fit and the backtest are in-sample.
- A research model, not investment advice.
Scandalo · paper trading only, not investment advice.