Teaching your agent

Market context

The six market features recorded at the moment of every decision.

At the moment of every decision, Kinfield records six numbers describing what the market was doing. This is what an agent “sees”: the context each trade is learned from, and the context object in the open dataset.

The six features

FieldShown asWhat it measures
ret5m5m momentumReturn over the last 5 minutes
ret15m15m momentumReturn over the last 15 minutes
ret60m1h momentumReturn over the last hour
vol15mVolatilityStandard deviation of the last 15 one-minute returns
rangePos60mRange positionWhere price sits in the last hour's range: 0 at the low, 1 at the high
volumeRatioVolume surgeAverage volume of the last 5 minutes against the hour's average

How they are computed

From one-minute candles, oldest to newest, with the last close replaced by the fill price. At least 61 candles are needed; Kinfield uses the last 90.

p          = the price the order met
close[k]   = the close k minutes ago (close[0] = p)

ret5m      = p / close[5]  - 1
ret15m     = p / close[15] - 1
ret60m     = p / close[60] - 1

r[i]       = close[i] / close[i+1] - 1        for the last 15 minutes
vol15m     = population standard deviation of r

hi, lo     = highest high and lowest low of the last 60 candles (and p)
rangePos60m = (p - lo) / (hi - lo)             0.5 if hi = lo

volumeRatio = mean volume of last 5 candles / mean volume of last 60

Returns are fractions, so 0.0042 means up 0.42%. For example, ret60m: -0.012 with rangePos60m: 0.08 describes a market down 1.2% on the hour, trading near its low.

When it is captured

  • Market orders: at the fill, from the same Hyperliquid price the order met.
  • Limit orders: when the order is placed, not when it fills, because placing it was the decision.
  • Signed in: Kinfield's server computes the features itself from Hyperliquid's candles. Nothing the browser sends is used.
  • As a guest: your browser computes them from the same candles.

If there isn't enough recent price history (for example, a market listed minutes ago), the context is left empty. The trade still counts on your record, but it can't teach your agent.

In the data

context, as published

{
  "ret5m": 0.00118,
  "ret15m": 0.00342,
  "ret60m": -0.00591,
  "vol15m": 0.00071,
  "rangePos60m": 0.62,
  "volumeRatio": 1.84
}

See the trade record for the full schema.