Probability
Every trading day the engine makes an up-or-down call on each symbol, then waits a day and marks its own homework. This page is the marked homework.
What this page is
Probability is a statistical model kept honest by a public scorecard. After each close it replays roughly the last 250 trading days for every symbol, makes a directional call for each of those days from that day's indicators and floor-sheet, and checks the call against what the price actually did on the next trading day.
The left table scores the whole market one day at a time: how many symbols were called up, how many were called down, how many actually rose and fell the next session, and the resulting accuracy. The right table breaks the selected day down symbol by symbol. The chart at the bottom plots one symbol's past calls against its actual next-day moves.
Because it is rebuilt only after the close, it never uses information from the future. Everything on the page has already been scored against reality.
Why it helps you
- Read it as a confidence weighting, not a signal. If the model has been running near 52% accuracy for a fortnight, then its lean on a stock you are already interested in deserves very little of your decision.
- The Log table's accuracy column shows that markets have moods. Long runs of high accuracy tend to be trending stretches where direction persists; the red days are the choppy ones, and those are also the days that hurt discretionary trading.
- The Track table is useful for the opposite of what people expect: finding the stocks where the model was confidently wrong. Those are often the names where something outside the indicators was driving the price.
- A next-day directional call has no bearing on a multi-year holding, and this page will not help you choose one. What it will do is show you, with real numbers, how weak short-horizon prediction is — which is itself an argument for holding.
- The Index and Sub-Index view is the one worth a glance. Accuracy on NEPSE and the sector indices reads as a rough measure of how orderly the market has been, which is context for whether to add now or wait.
When to use it, and where to find it
- After the close, once the daily rebuild has run. The status bar shows the date the data was last updated to.
- Once a week rather than every day — the useful signal here is the trend in accuracy, not a single day's reading.
- Before you lean on any model in this app, as a reality check on what these models are actually worth.
Try it — a short walkthrough
- 1Start with the accuracy columnScan the Accuracy badges down the Log table. Anything under 50% is worse than a coin flip and is shown red. Click the row of a day you want to examine.See this part in detail →
- 2Read the one-line summary for that dayThe System Log line restates the selected day in plain English — how many the model called each way, how many actually moved each way, and the hit rate.See this part in detail →
- 3Narrow the symbol listUse Show to pick Index and Sub-Index, or a single sector. The Track table on the right follows it.See this part in detail →
- 4Look for matching coloursIn the Track table, a row where Prob Change and Real Change are both green, or both red, is a call the model got right. Mixed colours are misses.See this part in detail →
- 5Chart one symbol's recordClick a row in the Track table. The chart plots that symbol's past calls as a solid line against its actual next-day moves as a dashed one — same side of zero means the call landed.See this part in detail →
Every part of the page
In the order you meet them on screen, each with a close-up of the real thing. The numbers match the dots on the figure above.
Date
Which past trading day you are examining. The Track table and the log line both follow this date.
How — Click a row in the Log table to jump to that day, or pick a date here directly.
- Only days the engine actually evaluated are available — it needs the FOLLOWING trading day to score the call.
- It defaults to the most recent day present in the data, not to today's calendar date.
Show
Narrows the Track table on the right to one group of symbols.
How — Pick "Index & Sub-Index" for NEPSE plus the 12 sector indices, or pick a sector to see only its ordinary stocks. The box is searchable.
- Remember that for an index the traded "volume" is turnover in rupees, while for an ordinary stock it is a share count.
System Log
A one-line, plain-English summary of the selected date, written directly from that day's real numbers — nothing is made up.
How — Read it as the headline for the day before you dig into the tables.
- It restates how many symbols the model called up and down, how many actually moved up and down the next session, and the hit-rate.
- It shows "No log entry found for this date" for a date the engine did not evaluate.
Log Part
One row per past trading day, scoring the whole market at once: how many stocks the model called up or down that day, versus how many actually moved up or down the next session.
How — Click a row to make that day the selected date. The Accuracy badge on the right is the quickest way to spot the good and bad days.
- + Count / − Count are the symbols predicted to rise / fall. Next + / Next − are the ones that actually closed higher / lower on the next day.
- + Diff / − Diff are actual minus predicted. Positive means more stocks rose than the model expected.
- Accuracy = of the symbols it gave a NON-NEUTRAL call on, the % where the predicted direction matched reality. Flat moves and neutral calls are excluded.
- A date is only listed once at least 30 symbols were evaluated on it. Below 50% accuracy shows red — worse than a coin flip.
Track Part
The same day broken down symbol by symbol: what the model leaned towards for each one, and what that stock actually did the next session.
How — Click a row to chart that symbol below. Matching colours — both green or both red — mean the call was right.
- IMPORTANT: "Prob Change" is shown with a % sign, but it is NOT a predicted price move. Its sign is the direction the model leaned and its size is a confidence score.
- "Real Change" IS a real percentage — the stock's actual close-to-close % move on the next trading day.
- The list is filtered by the "Show" selector in the top bar.
Projection
For the selected symbol, the model's past calls (solid blue "Prob") plotted against what the price actually did the next day (dashed green "Real"). It is a track record, not a forecast of tomorrow.
How — Look at which side of zero each line sits on: when both are on the same side, that call was right. Hover a point for the exact values.
- Prob above zero means the model expected a rise, below zero a fall — again, the height is confidence, not a predicted %.
- Every point is history. Nothing on this chart extends into the future.
Probability Model
The status bar names the model, the symbol currently charted, and the date the data was last updated to.
How — Open its ⓘ for the full list of inputs and the honest limits of the whole page in one place.
- If the "Updated" date is older than the last trading day, the daily rebuild has not run yet.
Reading the numbers
- Each call is built from that day's daily RSI(14) and the slope of its last five values, MACD(12,26,9) and its slope, intraday RSI from the 1-hour and 15-minute candles, whether volume rose, and the broker buy/sell concentration from that day's floor sheet.
- Prob Change is printed with a % sign but it is NOT a predicted price move. Its sign is the direction the model leaned; its size is a confidence score. Treat a +3.2% as "leaning up, fairly strongly", never as "expects a 3.2% rise".
- Real Change IS a true percentage — the stock's actual close-to-close move on the next trading day.
- Accuracy is measured only over symbols the model gave a non-neutral call on. Neutral calls and flat price moves are excluded from both the numerator and the denominator.
- + Count and − Count are symbols predicted to rise and fall. Next + and Next − are how many actually closed higher and lower. + Diff and − Diff are actual minus predicted, so a positive + Diff means more stocks rose than the model expected.
- A date only appears in the Log table once at least 30 symbols were evaluated on it, so thin days are dropped rather than shown with an unstable percentage.
- The date picker only offers days the engine actually scored — it needs the following trading day to exist before a call can be marked. It opens on the latest day present in the data, not on today's calendar date.
What it does not do
- Everything on this page is a description of the past. The model has already been told the answer for every row you can see, and nothing here extends into tomorrow. A high historical hit-rate is evidence that the inputs carried some information on those days; it is not a promise about the next one.
- Accuracy hovering near 50% is the normal state for next-day direction, and it means the model is close to worthless as a standalone signal. Read that plainly rather than hunting for the days it happened to be right.
- A directional call says nothing about size. The model can be right on direction for a week and still lose money for anyone trading it, because the wrong days move further than the right ones.
- The inputs are indicators and floor-sheet concentration only. News, earnings, corporate actions and anything else that actually moves a price are invisible to it.
- If the Updated date in the status bar is older than the last trading day, the daily rebuild has not run and you are reading stale scores.
- Days with fewer than 30 evaluated symbols are hidden rather than flagged, so a quiet market period can leave gaps in the Log table.
Questions people ask
No. The sign is the direction it leaned and the number is a confidence score that happens to be rendered with a percent sign. Only Real Change is an actual percentage move.
A call cannot be scored until the next trading day has closed. The picker only offers days that have already been marked.
Not broken — just close to a coin flip on that day, which is common for next-day direction. That is exactly what the page exists to show you.
Probability scores a single next-day directional call across the whole market. Trading Horizon replays a complete buy-and-sell rule on one stock and reports the money it would have made.